Computer Vision and Pattern Recognition (cs.CV)

Abstract: Human-Object Interaction (HOI) detection is a task to predict interactions between humans and objects in an image. In real-world scenarios, HOI detection models are required systematic generalization, i.e., generalization to novel combinations of objects and interactions, because it is highly probable that the train data only cover a limited portion of all possible combinations. However, to our knowledge, no open benchmark or existing work evaluates the systematic generalization in HOI detection. To address this issue, we created two new sets of HOI detection data splits named HICO-DET-SG and V-COCO-SG based on HICO-DET and V-COCO datasets. We evaluated representative HOI detection models on the new data splits and observed large degradation in the test performances compared to those on the original datasets. This result shows that systematic generalization is a challenging goal in HOI detection. We hope our new data splits encourage more research toward this goal.

Abstract: In this work, we introduce a novel deep learning architecture, Variable Length Embeddings (VLEs), an autoregressive model that can produce a latent representation composed of an arbitrary number of tokens. As a proof of concept, we demonstrate the capabilities of VLEs on tasks that involve reconstruction and image decomposition. We evaluate our experiments on a mix of the iNaturalist and ImageNet datasets and find that VLEs achieve comparable reconstruction results to a state of the art VAE, using less than a tenth of the parameters.

Abstract: This paper presents a generalized model for real-time detection of flying objects that can be used for transfer learning and further research, as well as a refined model that is ready for implementation. We achieve this by training our first generalized model on a data set containing 40 different classes of flying objects, forcing the model to extract abstract feature representations. We then perform transfer learning with these learned parameters on a data set more representative of real world environments (i.e., higher frequency of occlusion, small spatial sizes, rotations, etc.) to generate our refined model. Object detection of flying objects remains challenging due to large variance object spatial sizes/aspect ratios, rate of speed, occlusion, and clustered backgrounds. To address some of the presented challenges while simultaneously maximizing performance, we utilize the current state of the art single-shot detector, YOLOv8, in an attempt to find the best tradeoff between inference speed and mAP. While YOLOv8 is being regarded as the new state-of-the-art, an official paper has not been provided. Thus, we provide an in-depth explanation of the new architecture and functionality that YOLOv8 has adapted. Our final generalized model achieves an mAP50-95 of 0.685 and average inference speed on 1080p videos of 50 fps. Our final refined model maintains this inference speed and achieves an improved mAP50-95 of 0.835.

Abstract: In this work, we study self-supervised multiple object tracking without using any video-level association labels. We propose to cast the problem of multiple object tracking as learning the frame-wise associations between detections in consecutive frames. To this end, we propose differentiable soft object assignment for object association, making it possible to learn features tailored to object association with differentiable end-to-end training. With this training approach in hand, we develop an appearance-based model for learning instance-aware object features used to construct a cost matrix based on the pairwise distances between the object features. We train our model using temporal and multi-view data, where we obtain association pseudo-labels using optical flow and disparity information. Unlike most self-supervised tracking methods that rely on pretext tasks for learning the feature correspondences, our method is directly optimized for cross-object association in complex scenarios. As such, the proposed method offers a reidentification-based MOT approach that is robust to training hyperparameters and does not suffer from local minima, which are a challenge in self-supervised methods. We evaluate our proposed model on the KITTI, Waymo, nuScenes, and Argoverse datasets, consistently improving over other unsupervised methods ($7.8\%$ improvement in association accuracy on nuScenes).

Abstract: Adverse conditions typically suffer from stochastic hybrid weather degradations (e.g., rainy and hazy night), while existing image restoration algorithms envisage that weather degradations occur independently, thus may fail to handle real-world complicated scenarios. Besides, supervised training is not feasible due to the lack of comprehensive paired dataset to characterize hybrid conditions. To this end, we have advanced the forementioned limitations with two tactics: framework and data. On the one hand, we present a novel unified framework, dubbed RAHC, to Restore Arbitrary Hybrid adverse weather Conditions in one go, which can comfortably cope with hybrid scenarios with insufficient remaining background constituents and restore arbitrary hybrid conditions with a single trained model flexibly. On the other hand, we establish a new dataset, termed HAC, for learning and benchmarking arbitrary Hybrid Adverse Conditions restoration. HAC contains 31 scenarios composed of an arbitrary combination of five common weather, with a total of ~316K adverse-weather/clean pairs. As for fabrication, the training set is automatically generated by a dedicated AdverseGAN with no-frills labor, while the test set is manually modulated by experts for authoritative evaluation. Extensive experiments yield superior results and in particular establish new state-of-the-art results on both HAC and conventional datasets.

Abstract: This research focuses on the discovery and localization of hidden objects in the wild and serves unmanned systems. Through empirical analysis, infrared and visible image fusion (IVIF) enables hard-to-find objects apparent, whereas multimodal salient object detection (SOD) accurately delineates the precise spatial location of objects within the picture. Their common characteristic of seeking complementary cues from different source images motivates us to explore the collaborative relationship between Fusion and Salient object detection tasks on infrared and visible images via an Interactively Reinforced multi-task paradigm for the first time, termed IRFS. To the seamless bridge of multimodal image fusion and SOD tasks, we specifically develop a Feature Screening-based Fusion subnetwork (FSFNet) to screen out interfering features from source images, thereby preserving saliency-related features. After generating the fused image through FSFNet, it is then fed into the subsequent Fusion-Guided Cross-Complementary SOD subnetwork (FC$^2$Net) as the third modality to drive the precise prediction of the saliency map by leveraging the complementary information derived from the fused image. In addition, we develop an interactive loop learning strategy to achieve the mutual reinforcement of IVIF and SOD tasks with a shorter training period and fewer network parameters. Comprehensive experiment results demonstrate that the seamless bridge of IVIF and SOD mutually enhances their performance, and highlights their superiority.

Abstract: Video snapshot compressive imaging (SCI) uses a two-dimensional detector to capture consecutive video frames during a single exposure time. Following this, an efficient reconstruction algorithm needs to be designed to reconstruct the desired video frames. Although recent deep learning-based state-of-the-art (SOTA) reconstruction algorithms have achieved good results in most tasks, they still face the following challenges due to excessive model complexity and GPU memory limitations: 1) these models need high computational cost, and 2) they are usually unable to reconstruct large-scale video frames at high compression ratios. To address these issues, we develop an {\bf{\em efficient network}} for video SCI by using {\bf {\em dense connections and space-time factorization mechanism}} within a single residual block, dubbed {\bf \emph{EfficientSCI}}. The EfficientSCI network can well establish spatial-temporal correlation by using {\bf {\em convolution in the spatial domain and Transformer in the temporal domain}}, respectively. We are the first time to show that an UHD color video with high compression ratio can be reconstructed from a snapshot 2D measurement using a single end-to-end deep learning model with PSNR above 32 dB. Extensive results on both simulation and real data show that our method significantly outperforms all previous SOTA algorithms with better real-time performance. The code is at \url{https://github.com/ucaswangls/EfficientSCI.git}.

Abstract: Fine-grained classification is a challenging task that involves identifying subtle differences between objects within the same category. This task is particularly challenging in scenarios where data is scarce. Visual transformers (ViT) have recently emerged as a powerful tool for image classification, due to their ability to learn highly expressive representations of visual data using self-attention mechanisms. In this work, we explore Semi-ViT, a ViT model fine tuned using semi-supervised learning techniques, suitable for situations where we have lack of annotated data. This is particularly common in e-commerce, where images are readily available but labels are noisy, nonexistent, or expensive to obtain. Our results demonstrate that Semi-ViT outperforms traditional convolutional neural networks (CNN) and ViTs, even when fine-tuned with limited annotated data. These findings indicate that Semi-ViTs hold significant promise for applications that require precise and fine-grained classification of visual data.

Abstract: Colonoscopy is the most widely used medical technique for preventing Colorectal Cancer, by detecting and removing polyps before they become malignant. Recent studies show that around one quarter of the existing polyps are routinely missed. While some of these do appear in the endoscopist's field of view, others are missed due to a partial coverage of the colon. The task of detecting and marking unseen regions of the colon has been addressed in recent work, where the common approach is based on dense 3D reconstruction, which proves to be challenging due to lack of 3D ground truth and periods with poor visual content. In this paper we propose a novel and complementary method to detect deficient local coverage in real-time for video segments where a reliable 3D reconstruction is impossible. Our method aims to identify skips along the colon caused by a drifted position of the endoscope during poor visibility time intervals. The proposed solution consists of two phases. During the first, time segments with good visibility of the colon and gaps between them are identified. During the second phase, a trained model operates on each gap, answering the question: Do you observe the same scene before and after the gap? If the answer is negative, the endoscopist is alerted and can be directed to the appropriate area in real-time. The second phase model is trained using a contrastive loss based on auto-generated examples. Our method evaluation on a dataset of 250 procedures annotated by trained physicians provides sensitivity of 0.75 with specificity of 0.9.

Abstract: Recovering noise-covered details from low-light images is challenging, and the results given by previous methods leave room for improvement. Recent diffusion models show realistic and detailed image generation through a sequence of denoising refinements and motivate us to introduce them to low-light image enhancement for recovering realistic details. However, we found two problems when doing this, i.e., 1) diffusion models keep constant resolution in one reverse process, which limits the speed; 2) diffusion models sometimes result in global degradation (e.g., RGB shift). To address the above problems, this paper proposes a Pyramid Diffusion model (PyDiff) for low-light image enhancement. PyDiff uses a novel pyramid diffusion method to perform sampling in a pyramid resolution style (i.e., progressively increasing resolution in one reverse process). Pyramid diffusion makes PyDiff much faster than vanilla diffusion models and introduces no performance degradation. Furthermore, PyDiff uses a global corrector to alleviate the global degradation that may occur in the reverse process, significantly improving the performance and making the training of diffusion models easier with little additional computational consumption. Extensive experiments on popular benchmarks show that PyDiff achieves superior performance and efficiency. Moreover, PyDiff can generalize well to unseen noise and illumination distributions.

Abstract: We propose a novel visual SLAM method that integrates text objects tightly by treating them as semantic features via fully exploring their geometric and semantic prior. The text object is modeled as a texture-rich planar patch whose semantic meaning is extracted and updated on the fly for better data association. With the full exploration of locally planar characteristics and semantic meaning of text objects, the SLAM system becomes more accurate and robust even under challenging conditions such as image blurring, large viewpoint changes, and significant illumination variations (day and night). We tested our method in various scenes with the ground truth data. The results show that integrating texture features leads to a more superior SLAM system that can match images across day and night. The reconstructed semantic 3D text map could be useful for navigation and scene understanding in robotic and mixed reality applications. Our project page: https://github.com/SJTU-ViSYS/TextSLAM .

Abstract: In implant prosthesis treatment, the design of surgical guide requires lots of manual labors and is prone to subjective variations. When deep learning based methods has started to be applied to address this problem, the space between teeth are various and some of them might present similar texture characteristic with the actual implant region. Both problems make a big challenge for the implant position prediction. In this paper, we develop a two-stream implant position regression framework (TSIPR), which consists of an implant region detector (IRD) and a multi-scale patch embedding regression network (MSPENet), to address this issue. For the training of IRD, we extend the original annotation to provide additional supervisory information, which contains much more rich characteristic and do not introduce extra labeling costs. A multi-scale patch embedding module is designed for the MSPENet to adaptively extract features from the images with various tooth spacing. The global-local feature interaction block is designed to build the encoder of MSPENet, which combines the transformer and convolution for enriched feature representation. During inference, the RoI mask extracted from the IRD is used to refine the prediction results of the MSPENet. Extensive experiments on a dental implant dataset through five-fold cross-validation demonstrated that the proposed TSIPR achieves superior performance than existing methods.

Abstract: Video question answering aims at answering a question about the video content by reasoning the alignment semantics within them. However, since relying heavily on human instructions, i.e., annotations or priors, current contrastive learning-based VideoQA methods remains challenging to perform fine-grained visual-linguistic alignments. In this work, we innovatively resort to game theory, which can simulate complicated relationships among multiple players with specific interaction strategies, e.g., video, question, and answer as ternary players, to achieve fine-grained alignment for VideoQA task. Specifically, we carefully design a VideoQA-specific interaction strategy to tailor the characteristics of VideoQA, which can mathematically generate the fine-grained visual-linguistic alignment label without label-intensive efforts. Our TG-VQA outperforms existing state-of-the-art by a large margin (more than 5%) on long-term and short-term VideoQA datasets, verifying its effectiveness and generalization ability. Thanks to the guidance of game-theoretic interaction, our model impressively convergences well on limited data (${10}^4 ~videos$), surpassing most of those pre-trained on large-scale data ($10^7~videos$).

Abstract: Oriented object detection has been developed rapidly in the past few years, where rotation equivariant is crucial for detectors to predict rotated bounding boxes. It is expected that the prediction can maintain the corresponding rotation when objects rotate, but severe mutational in angular prediction is sometimes observed when objects rotate near the boundary angle, which is well-known boundary discontinuity problem. The problem has been long believed to be caused by the sharp loss increase at the angular boundary during training, and widely used IoU-like loss generally deal with this problem by loss-smoothing. However, we experimentally find that even state-of-the-art IoU-like methods do not actually solve the problem. On further analysis, we find the essential cause of the problem lies at discontinuous angular ground-truth(box), not just discontinuous loss. There always exists an irreparable gap between continuous model ouput and discontinuous angular ground-truth, so angular prediction near the breakpoints becomes highly unstable, which cannot be eliminated just by loss-smoothing in IoU-like methods. To thoroughly solve this problem, we propose a simple and effective Angle Correct Module (ACM) based on polar coordinate decomposition. ACM can be easily plugged into the workflow of oriented object detectors to repair angular prediction. It converts the smooth value of the model output into sawtooth angular value, and then IoU-like loss can fully release their potential. Extensive experiments on multiple datasets show that whether Gaussian-based or SkewIoU methods are improved to the same performance of AP50 and AP75 with the enhancement of ACM.

Abstract: In many machine learning applications, labeling datasets can be an arduous and time-consuming task. Although research has shown that semi-supervised learning techniques can achieve high accuracy with very few labels within the field of computer vision, little attention has been given to how images within a dataset should be selected for labeling. In this paper, we propose a novel approach based on well-established self-supervised learning, clustering, and manifold learning techniques that address this challenge of selecting an informative image subset to label in the first instance, which is known as the cold-start or unsupervised selective labelling problem. We test our approach using several publicly available datasets, namely CIFAR10, Imagenette, DeepWeeds, and EuroSAT, and observe improved performance with both supervised and semi-supervised learning strategies when our label selection strategy is used, in comparison to random sampling. We also obtain superior performance for the datasets considered with a much simpler approach compared to other methods in the literature.

Abstract: Constructing structural brain networks using T1-weighted magnetic resonance imaging (T1-MRI) presents a significant challenge due to the lack of direct regional connectivity information. Current methods with T1-MRI rely on predefined regions or isolated pretrained location modules to obtain atrophic regions, which neglects individual specificity. Besides, existing methods capture global structural context only on the whole-image-level, which weaken correlation between regions and the hierarchical distribution nature of brain connectivity.We hereby propose a novel dynamic structural brain network construction method based on T1-MRI, which can dynamically localize critical regions and constrain the hierarchical distribution among them for constructing dynamic structural brain network. Specifically, we first cluster spatially-correlated channel and generate several critical brain regions as prototypes. Further, we introduce a contrastive loss function to constrain the prototypes distribution, which embed the hierarchical brain semantic structure into the latent space. Self-attention and GCN are then used to dynamically construct hierarchical correlations of critical regions for brain network and explore the correlation, respectively. Our method is evaluated on ADNI-1 and ADNI-2 databases for mild cognitive impairment (MCI) conversion prediction, and acheive the state-of-the-art (SOTA) performance. Our source code is available at http://github.com/*******.

Abstract: In the field of deep learning applied to face recognition, securing large-scale, high-quality datasets is vital for attaining precise and reliable results. However, amassing significant volumes of high-quality real data faces hurdles such as time limitations, financial burdens, and privacy issues. Furthermore, prevalent datasets are often impaired by racial biases and annotation inaccuracies. In this paper, we underscore the promising application of synthetic data, generated through rendering digital faces via our computer graphics pipeline, in achieving competitive results with the state-of-the-art on synthetic data across multiple benchmark datasets. By finetuning the model,we obtain results that rival those achieved when training with hundreds of thousands of real images (98.7% on LFW [1]). We further investigate the contribution of adding intra-class variance factors (e.g., makeup, accessories, haircuts) on model performance. Finally, we reveal the sensitivity of pre-trained face recognition models to alternating specific parts of the face by leveraging the granular control capability in our platform.

Abstract: Anomaly detection in temporal data from sensors under aviation scenarios is a practical but challenging task: 1) long temporal data is difficult to extract contextual information with temporal correlation; 2) the anomalous data are rare in time series, causing normal/abnormal imbalance in anomaly detection, making the detector classification degenerate or even fail. To remedy the aforementioned problems, we propose a Graphical Temporal Data Analysis (GTDA) framework. It consists three modules, named Series-to-Image (S2I), Cluster-based Resampling Approach using Euclidean Distance (CRD) and Variance-Based Loss (VBL). Specifically, for better extracts global information in temporal data from sensors, S2I converts the data to curve images to demonstrate abnormalities in data changes. CRD and VBL balance the classification to mitigate the unequal distribution of classes. CRD extracts minority samples with similar features to majority samples by clustering and over-samples them. And VBL fine-tunes the decision boundary by balancing the fitting degree of the network to each class. Ablation experiments on the Flights dataset indicate the effectiveness of CRD and VBL on precision and recall, respectively. Extensive experiments demonstrate the synergistic advantages of CRD and VBL on F1-score on Flights and three other temporal datasets.

Abstract: The growing demand for precision agriculture necessitates efficient and accurate crop-weed recognition and classification systems. Current datasets often lack the sample size, diversity, and hierarchical structure needed to develop robust deep learning models for discriminating crops and weeds in agricultural fields. Moreover, the similar external structure and phenomics of crops and weeds complicate recognition tasks. To address these issues, we present the CWD30 dataset, a large-scale, diverse, holistic, and hierarchical dataset tailored for crop-weed recognition tasks in precision agriculture. CWD30 comprises over 219,770 high-resolution images of 20 weed species and 10 crop species, encompassing various growth stages, multiple viewing angles, and environmental conditions. The images were collected from diverse agricultural fields across different geographic locations and seasons, ensuring a representative dataset. The dataset's hierarchical taxonomy enables fine-grained classification and facilitates the development of more accurate, robust, and generalizable deep learning models. We conduct extensive baseline experiments to validate the efficacy of the CWD30 dataset. Our experiments reveal that the dataset poses significant challenges due to intra-class variations, inter-class similarities, and data imbalance. Additionally, we demonstrate that minor training modifications like using CWD30 pretrained backbones can significantly enhance model performance and reduce convergence time, saving training resources on several downstream tasks. These challenges provide valuable insights and opportunities for future research in crop-weed recognition. We believe that the CWD30 dataset will serve as a benchmark for evaluating crop-weed recognition algorithms, promoting advancements in precision agriculture, and fostering collaboration among researchers in the field.

Abstract: Colonoscopy is the standard of care technique for detecting and removing polyps for the prevention of colorectal cancer. Nevertheless, gastroenterologists (GI) routinely miss approximately 25% of polyps during colonoscopies. These misses are highly operator dependent, influenced by the physician skills, experience, vigilance, and fatigue. Standard quality metrics, such as Withdrawal Time or Cecal Intubation Rate, have been shown to be well correlated with Adenoma Detection Rate (ADR). However, those metrics are limited in their ability to assess the quality of a specific procedure, and they do not address quality aspects related to the style or technique of the examination. In this work we design novel online and offline quality metrics, based on visual appearance quality criteria learned by an ML model in an unsupervised way. Furthermore, we evaluate the likelihood of detecting an existing polyp as a function of quality and use it to demonstrate high correlation of the proposed metric to polyp detection sensitivity. The proposed online quality metric can be used to provide real time quality feedback to the performing GI. By integrating the local metric over the withdrawal phase, we build a global, offline quality metric, which is shown to be highly correlated to the standard Polyp Per Colonoscopy (PPC) quality metric.

Abstract: Filter-decomposition-based group-equivariant convolutional neural networks (G-CNN) have been demonstrated to increase CNN's data efficiency and contribute to better interpretability and controllability of CNN models. However, so far filter-decomposition-based affine G-CNN methods rely on parameter sharing for achieving high parameter efficiency and suffer from a heavy computational burden. They also use a limited number of transformations and in particular ignore the shear transform in the application. In this paper, we address these problems by emphasizing the importance of the diversity of transformations. We propose a flexible and efficient strategy based on weighted filter-wise Monte Carlo sampling. In addition, we introduce shear equivariant CNN to address the highly sparse representations of natural images. We demonstrate that the proposed methods are intrinsically an efficient generalization of traditional CNNs, and we explain the advantage of bottleneck architectures used in the existing state-of-the-art CNN models such as ResNet, ResNext, and ConvNeXt from the group-equivariant perspective. Experiments on image classification and image denoising tasks show that with a set of suitable filter basis, our methods achieve superior performance to standard CNN with high data efficiency. The code will be available at https://github.com/ZhaoWenzhao/MCG_CNN.

Abstract: Acquiring and annotating suitable datasets for training deep learning models is challenging. This often results in tedious and time-consuming efforts that can hinder research progress. However, generative models have emerged as a promising solution for generating synthetic datasets that can replace or augment real-world data. Despite this, the effectiveness of synthetic data is limited by their inability to fully capture the complexity and diversity of real-world data. To address this issue, we explore the use of Generative Adversarial Networks to generate synthetic datasets for training classifiers that are subsequently evaluated on real-world images. To improve the quality and diversity of the synthetic dataset, we propose three novel post-processing techniques: Dynamic Sample Filtering, Dynamic Dataset Recycle, and Expansion Trick. In addition, we introduce a pipeline called Gap Filler (GaFi), which applies these techniques in an optimal and coordinated manner to maximise classification accuracy on real-world data. Our experiments show that GaFi effectively reduces the gap with real-accuracy scores to an error of 2.03%, 1.78%, and 3.99% on the Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets, respectively. These results represent a new state of the art in Classification Accuracy Score and highlight the effectiveness of post-processing techniques in improving the quality of synthetic datasets.

Abstract: With the continue development of Convolutional Neural Networks (CNNs), there is a growing concern regarding representations that they encode internally. Analyzing these internal representations is referred to as model interpretation. While the task of model explanation, justifying the predictions of such models, has been studied extensively; the task of model interpretation has received less attention. The aim of this paper is to propose a framework for the study of interpretation methods designed for CNN models trained from visual data. More specifically, we first specify the difference between the interpretation and explanation tasks which are often considered the same in the literature. Then, we define a set of six specific factors that can be used to characterize interpretation methods. Third, based on the previous factors, we propose a framework for the positioning of interpretation methods. Our framework highlights that just a very small amount of the suggested factors, and combinations thereof, have been actually studied. Consequently, leaving significant areas unexplored. Following the proposed framework, we discuss existing interpretation methods and give some attention to the evaluation protocols followed to validate them. Finally, the paper highlights capabilities of the methods in producing feedback for enabling interpretation and proposes possible research problems arising from the framework.

Abstract: Uncertainty quantification for inverse problems in imaging has drawn much attention lately. Existing approaches towards this task define uncertainty regions based on probable values per pixel, while ignoring spatial correlations within the image, resulting in an exaggerated volume of uncertainty. In this paper, we propose PUQ (Principal Uncertainty Quantification) -- a novel definition and corresponding analysis of uncertainty regions that takes into account spatial relationships within the image, thus providing reduced volume regions. Using recent advancements in stochastic generative models, we derive uncertainty intervals around principal components of the empirical posterior distribution, forming an ambiguity region that guarantees the inclusion of true unseen values with a user confidence probability. To improve computational efficiency and interpretability, we also guarantee the recovery of true unseen values using only a few principal directions, resulting in ultimately more informative uncertainty regions. Our approach is verified through experiments on image colorization, super-resolution, and inpainting; its effectiveness is shown through comparison to baseline methods, demonstrating significantly tighter uncertainty regions.

Abstract: The goal of a speech-to-image transform is to produce a photo-realistic picture directly from a speech signal. Recently, various studies have focused on this task and have achieved promising performance. However, current speech-to-image approaches are based on a stacked modular framework that suffers from three vital issues: 1) Training separate networks is time-consuming as well as inefficient and the convergence of the final generative model strongly depends on the previous generators; 2) The quality of precursor images is ignored by this architecture; 3) Multiple discriminator networks are required to be trained. To this end, we propose an efficient and effective single-stage framework called Fusion-S2iGan to yield perceptually plausible and semantically consistent image samples on the basis of given spoken descriptions. Fusion-S2iGan introduces a visual+speech fusion module (VSFM), constructed with a pixel-attention module (PAM), a speech-modulation module (SMM) and a weighted-fusion module (WFM), to inject the speech embedding from a speech encoder into the generator while improving the quality of synthesized pictures. Fusion-S2iGan spreads the bimodal information over all layers of the generator network to reinforce the visual feature maps at various hierarchical levels in the architecture. We conduct a series of experiments on four benchmark data sets, i.e., CUB birds, Oxford-102, Flickr8k and Places-subset. The experimental results demonstrate the superiority of the presented Fusion-S2iGan compared to the state-of-the-art models with a multi-stage architecture and a performance level that is close to traditional text-to-image approaches.

Abstract: This paper presents a fully automatic registration method of dental cone-beam computed tomography (CBCT) and face scan data. It can be used for a digital platform of 3D jaw-teeth-face models in a variety of applications, including 3D digital treatment planning and orthognathic surgery. Difficulties in accurately merging facial scans and CBCT images are due to the different image acquisition methods and limited area of correspondence between the two facial surfaces. In addition, it is difficult to use machine learning techniques because they use face-related 3D medical data with radiation exposure, which are difficult to obtain for training. The proposed method addresses these problems by reusing an existing machine-learning-based 2D landmark detection algorithm in an open-source library and developing a novel mathematical algorithm that identifies paired 3D landmarks from knowledge of the corresponding 2D landmarks. A main contribution of this study is that the proposed method does not require annotated training data of facial landmarks because it uses a pre-trained facial landmark detection algorithm that is known to be robust and generalized to various 2D face image models. Note that this reduces a 3D landmark detection problem to a 2D problem of identifying the corresponding landmarks on two 2D projection images generated from two different projection angles. Here, the 3D landmarks for registration were selected from the sub-surfaces with the least geometric change under the CBCT and face scan environments. For the final fine-tuning of the registration, the Iterative Closest Point method was applied, which utilizes geometrical information around the 3D landmarks. The experimental results show that the proposed method achieved an averaged surface distance error of 0.74 mm for three pairs of CBCT and face scan datasets.

Abstract: Brain signal visualization has emerged as an active research area, serving as a critical interface between the human visual system and computer vision models. Although diffusion models have shown promise in analyzing functional magnetic resonance imaging (fMRI) data, including reconstructing high-quality images consistent with original visual stimuli, their accuracy in extracting semantic and silhouette information from brain signals remains limited. In this regard, we propose a novel approach, referred to as Controllable Mind Visual Diffusion Model (CMVDM). CMVDM extracts semantic and silhouette information from fMRI data using attribute alignment and assistant networks. Additionally, a residual block is incorporated to capture information beyond semantic and silhouette features. We then leverage a control model to fully exploit the extracted information for image synthesis, resulting in generated images that closely resemble the visual stimuli in terms of semantics and silhouette. Through extensive experimentation, we demonstrate that CMVDM outperforms existing state-of-the-art methods both qualitatively and quantitatively.

Abstract: Currently, image-denoising methods based on deep learning cannot adequately reconcile contextual semantic information and spatial details. To take these information optimizations into consideration, in this paper, we propose a Context-Space Progressive Collaborative Network (CS-PCN) for image denoising. CS-PCN is a multi-stage hierarchical architecture composed of a context mining siamese sub-network (CM2S) and a space synthesis sub-network (3S). CM2S aims at extracting rich multi-scale contextual information by sequentially connecting multi-layer feature processors (MLFP) for semantic information pre-processing, attention encoder-decoders (AED) for multi-scale information, and multi-conv attention controllers (MCAC) for supervised feature fusion. 3S parallels MLFP and a single-scale cascading block to learn image details, which not only maintains the contextual information but also emphasizes the complementary spatial ones. Experimental results show that CS-PCN achieves significant performance improvement in synthetic and real-world noise removal.

Abstract: Video frame interpolation (VFI) increases the video frame rate by inserting a reconstruction frame into two consecutive frames. Due to the limitation of the fixed frame rate of ordinary camera, the frame-only video frame interpolation methods inevitably lose the dynamics in the interval between consecutive frames. In order to compensate for the lack of inter-frame information, motion models are often used, but those models cannot account for the real motions. Event cameras are bio-inspired vision sensor, each pixel of which independently perceives and encodes relative changes in light intensity. Event cameras output sparse, asynchronous streams of events instead of frames, with advantages of high temporal resolution, high dynamics, and low power consumption. An event is usually expressed as a tuple e=(x,y,p,t), which means that at timestamp t, an event with polarity is generated at the pixel (x,y). Positive polarity indicates that the change of light intensity from week to strong is beyond the threshold, while negative polarity is just the opposite. Because an event camera has high temporal resolution up to microseconds, it can capture complete changes or motion between frames. The event flow is the embodiment of inter-frame changes. Therefore, the optical flow estimated from the events does not require any motion model to be fitted, which can be inherently nonlinear. Since events lack intensity information, frame-based optical flow is complementary to event-based optical flow. By combining these two kinds of optical flow, more accurate estimation results can be obtained. Meanwhile, it is possible to reconstruct high-quality keyframes at any timestamp, since real inter-frame dynamics are captured.

Abstract: In this work, we study the problem of object re-identification (ReID) in a 3D multi-object tracking (MOT) context, by learning to match pairs of objects from cropped (e.g., using their predicted 3D bounding boxes) point cloud observations. We are not concerned with SOTA performance for 3D MOT, however. Instead, we seek to answer the following question: In a realistic tracking by-detection context, how does object ReID from point clouds perform relative to ReID from images? To enable such a study, we propose a lightweight matching head that can be concatenated to any set or sequence processing backbone (e.g., PointNet or ViT), creating a family of comparable object ReID networks for both modalities. Run in siamese style, our proposed point-cloud ReID networks can make thousands of pairwise comparisons in real-time (10 hz). Our findings demonstrate that their performance increases with higher sensor resolution and approaches that of image ReID when observations are sufficiently dense. Additionally, we investigate our network's ability to enhance 3D multi-object tracking (MOT), showing that our point-cloud ReID networks can successfully re-identify objects which led a strong motion-based tracker into error. To our knowledge, we are the first to study real-time object re-identification from point clouds in a 3D multi-object tracking context.

Abstract: Low-light situations severely restrict the pursuit of aesthetic quality in consumer photography. Although many efforts are devoted to designing heuristics, it is generally mired in a shallow spiral of tedium, such as piling up complex network architectures and empirical strategies. How to delve into the essential physical principles of illumination compensation has been neglected. Following the way of simplifying the complexity, this paper innovatively proposes a simple and efficient Noise-Free Illumination Interpolator (NFI$_2$). According to the constraint principle of illuminance and reflectance within a limited dynamic range, as a prior knowledge in the recovery process, we construct a learnable illuminance interpolator and thereby compensating for non-uniform lighting. With the intention of adapting denoising without annotated data, we design a self-calibrated denoiser with the intrinsic image properties to acquire noise-free low-light images. Starting from the properties of natural image manifolds, a self-regularized recovery loss is introduced as a way to encourage more natural and realistic reflectance map. The model architecture and training losses, guided by prior knowledge, complement and benefit each other, forming a powerful unsupervised leaning framework. Comprehensive experiments demonstrate that the proposed algorithm produces competitive qualitative and quantitative results while maintaining favorable generalization capability in unknown real-world scenarios.

Abstract: Copy-move forgery detection is a crucial research area within digital image forensics, as it focuses on identifying instances where objects in an image are duplicated and placed in different locations. The detection of such forgeries is particularly important in contexts where they can be exploited for malicious purposes. Recent years have witnessed an increased interest in distinguishing between the original and duplicated objects in copy-move forgeries, accompanied by the development of larger-scale datasets to facilitate this task. However, existing approaches to copy-move forgery detection and source/target differentiation often involve two separate steps or the design of individual end-to-end networks for each task. In this paper, we propose an innovative method that employs the transformer architecture in an end-to-end deep neural network. Our method aims to detect instances of copy-move forgery while simultaneously localizing the source and target regions. By utilizing this approach, we address the challenges posed by multi-object copy-move scenarios and report if there is a balance between the detection and differentiation tasks. To evaluate the performance of our proposed network, we conducted experiments on two publicly available copy-move datasets. The results and analysis aims to show the potential significance of our focus in balancing detection and distinguishment result and transferring the trained model in different datasets in the field.

Abstract: In recent years, the agricultural industry has witnessed significant advancements in artificial intelligence (AI), particularly with the development of large-scale foundational models. Among these foundation models, the Segment Anything Model (SAM), introduced by Meta AI Research, stands out as a groundbreaking solution for object segmentation tasks. While SAM has shown success in various agricultural applications, its potential in the poultry industry, specifically in the context of cage-free hens, remains relatively unexplored. This study aims to assess the zero-shot segmentation performance of SAM on representative chicken segmentation tasks, including part-based segmentation and the use of infrared thermal images, and to explore chicken-tracking tasks by using SAM as a segmentation tool. The results demonstrate SAM's superior performance compared to SegFormer and SETR in both whole and part-based chicken segmentation. SAM-based object tracking also provides valuable data on the behavior and movement patterns of broiler birds. The findings of this study contribute to a better understanding of SAM's potential in poultry science and lay the foundation for future advancements in chicken segmentation and tracking.

Abstract: Attribute-specific fashion retrieval (ASFR) is a challenging information retrieval task, which has attracted increasing attention in recent years. Different from traditional fashion retrieval which mainly focuses on optimizing holistic similarity, the ASFR task concentrates on attribute-specific similarity, resulting in more fine-grained and interpretable retrieval results. As the attribute-specific similarity typically corresponds to the specific subtle regions of images, we propose a Region-to-Patch Framework (RPF) that consists of a region-aware branch and a patch-aware branch to extract fine-grained attribute-related visual features for precise retrieval in a coarse-to-fine manner. In particular, the region-aware branch is first to be utilized to locate the potential regions related to the semantic of the given attribute. Then, considering that the located region is coarse and still contains the background visual contents, the patch-aware branch is proposed to capture patch-wise attribute-related details from the previous amplified region. Such a hybrid architecture strikes a proper balance between region localization and feature extraction. Besides, different from previous works that solely focus on discriminating the attribute-relevant foreground visual features, we argue that the attribute-irrelevant background features are also crucial for distinguishing the detailed visual contexts in a contrastive manner. Therefore, a novel E-InfoNCE loss based on the foreground and background representations is further proposed to improve the discrimination of attribute-specific representation. Extensive experiments on three datasets demonstrate the effectiveness of our proposed framework, and also show a decent generalization of our RPF on out-of-domain fashion images. Our source code is available at https://github.com/HuiGuanLab/RPF.

Abstract: EXplainable AI (XAI) is an essential topic to improve human understanding of deep neural networks (DNNs) given their black-box internals. For computer vision tasks, mainstream pixel-based XAI methods explain DNN decisions by identifying important pixels, and emerging concept-based XAI explore forming explanations with concepts (e.g., a head in an image). However, pixels are generally hard to interpret and sensitive to the imprecision of XAI methods, whereas "concepts" in prior works require human annotation or are limited to pre-defined concept sets. On the other hand, driven by large-scale pre-training, Segment Anything Model (SAM) has been demonstrated as a powerful and promotable framework for performing precise and comprehensive instance segmentation, enabling automatic preparation of concept sets from a given image. This paper for the first time explores using SAM to augment concept-based XAI. We offer an effective and flexible concept-based explanation method, namely Explain Any Concept (EAC), which explains DNN decisions with any concept. While SAM is highly effective and offers an "out-of-the-box" instance segmentation, it is costly when being integrated into defacto XAI pipelines. We thus propose a lightweight per-input equivalent (PIE) scheme, enabling efficient explanation with a surrogate model. Our evaluation over two popular datasets (ImageNet and COCO) illustrate the highly encouraging performance of EAC over commonly-used XAI methods.

Abstract: Mixup is a widely adopted strategy for training deep networks, where additional samples are augmented by interpolating inputs and labels of training pairs. Mixup has shown to improve classification performance, network calibration, and out-of-distribution generalisation. While effective, a cornerstone of Mixup, namely that networks learn linear behaviour patterns between classes, is only indirectly enforced since the output interpolation is performed at the probability level. This paper seeks to address this limitation by mixing the classifiers directly instead of mixing the labels for each mixed pair. We propose to define the target of each augmented sample as a uniquely new classifier, whose parameters are a linear interpolation of the classifier vectors of the input pair. The space of all possible classifiers is continuous and spans all interpolations between classifier pairs. To make optimisation tractable, we propose a dual-contrastive Infinite Class Mixup loss, where we contrast the classifier of a mixed pair to both the classifiers and the predicted outputs of other mixed pairs in a batch. Infinite Class Mixup is generic in nature and applies to many variants of Mixup. Empirically, we show that it outperforms standard Mixup and variants such as RegMixup and Remix on balanced, long-tailed, and data-constrained benchmarks, highlighting its broad applicability.

Abstract: Existing deep learning models for hyperspectral image (HSI) reconstruction achieve good performance but require powerful hardwares with enormous memory and computational resources. Consequently, these methods can hardly be deployed on resource-limited mobile devices. In this paper, we propose a novel method, Binarized Spectral-Redistribution Network (BiSRNet), for efficient and practical HSI restoration from compressed measurement in snapshot compressive imaging (SCI) systems. Firstly, we redesign a compact and easy-to-deploy base model to be binarized. Then we present the basic unit, Binarized Spectral-Redistribution Convolution (BiSR-Conv). BiSR-Conv can adaptively redistribute the HSI representations before binarizing activation and uses a scalable hyperbolic tangent function to closer approximate the Sign function in backpropagation. Based on our BiSR-Conv, we customize four binarized convolutional modules to address the dimension mismatch and propagate full-precision information throughout the whole network. Finally, our BiSRNet is derived by using the proposed techniques to binarize the base model. Comprehensive quantitative and qualitative experiments manifest that our proposed BiSRNet outperforms state-of-the-art binarization methods and achieves comparable performance with full-precision algorithms. Code and models will be released at https://github.com/caiyuanhao1998/BiSCI and https://github.com/caiyuanhao1998/MST

Abstract: Site-specific weed control (SSWC) can provide considerable reductions in weed control costs and herbicide usage. Despite the promise of machine vision for SSWC systems and the importance of ground speed in weed control efficacy, there has been little investigation of the role of ground speed and camera characteristics on weed detection performance. Here, we compare the performance of four camera-software combinations using the open-source OpenWeedLocator platform - (1) default settings on a Raspberry Pi HQ camera, (2) optimised software settings on a HQ camera, (3) optimised software settings on the Raspberry Pi v2 camera, and (4) a global shutter Arducam AR0234 camera - at speeds ranging from 5 km h$^{-1}$ to 30 km h$^{-1}$. A combined excess green (ExG) and hue, saturation, value (HSV) thresholding algorithm was used for testing under fallow conditions using tillage radish (Raphanus sativus) and forage oats (Avena sativa) as representative broadleaf and grass weeds, respectively. ARD demonstrated the highest recall among camera systems, with up to 95.7% of weeds detected at 5 km h$^{-1}$ and 85.7% at 30 km h$^{-1}$. HQ1 and V2 cameras had the lowest recall of 31.1% and 26.0% at 30 km h$^{-1}$, respectively. All cameras experienced a decrease in recall as speed increased. The highest rate of decrease was observed for HQ1 with 1.12% and 0.90% reductions in recall for every km h$^{-1}$ increase in speed for tillage radish and forage oats, respectively. Detection of the grassy forage oats was worse (P<0.05) than the broadleaved tillage radish for all cameras. Despite the variations in recall, HQ1, HQ2, and V2 maintained near-perfect precision at all tested speeds. The variable effect of ground speed and camera system on detection performance of grass and broadleaf weeds, indicates that careful hardware and software considerations must be made when developing SSWC systems.

Abstract: We apply convolutional neural networks (CNN) to the problem of image orientation detection in the context of determining the correct orientation (from 0, 90, 180, and 270 degrees) of a consumer photo. The problem is especially important for digitazing analog photographs. We substantially improve on the published state of the art in terms of the performance on one of the standard datasets, and test our system on a more difficult large dataset of consumer photos. We use Guided Backpropagation to obtain insights into how our CNN detects photo orientation, and to explain its mistakes.

Abstract: The core of Multi-view Stereo(MVS) is the matching process among reference and source pixels. Cost aggregation plays a significant role in this process, while previous methods focus on handling it via CNNs. This may inherit the natural limitation of CNNs that fail to discriminate repetitive or incorrect matches due to limited local receptive fields. To handle the issue, we aim to involve Transformer into cost aggregation. However, another problem may occur due to the quadratically growing computational complexity caused by Transformer, resulting in memory overflow and inference latency. In this paper, we overcome these limits with an efficient Transformer-based cost aggregation network, namely CostFormer. The Residual Depth-Aware Cost Transformer(RDACT) is proposed to aggregate long-range features on cost volume via self-attention mechanisms along the depth and spatial dimensions. Furthermore, Residual Regression Transformer(RRT) is proposed to enhance spatial attention. The proposed method is a universal plug-in to improve learning-based MVS methods.

Abstract: Single-Photon Emission Computed Tomography (SPECT) is widely applied for the diagnosis of ischemic heart diseases. Low-dose (LD) SPECT aims to minimize radiation exposure but leads to increased image noise. Limited-angle (LA) SPECT enables faster scanning and reduced hardware costs but results in lower reconstruction accuracy. Additionally, computed tomography (CT)-derived attenuation maps ($\mu$-maps) are commonly used for SPECT attenuation correction (AC), but it will cause extra radiation exposure and SPECT-CT misalignments. In addition, the majority of SPECT scanners in the market are not hybrid SPECT/CT scanners. Although various deep learning methods have been introduced to separately address these limitations, the solution for simultaneously addressing these challenges still remains highly under-explored and challenging. To this end, we propose a Cross-domain Iterative Network (CDI-Net) for simultaneous denoising, LA reconstruction, and CT-free AC in cardiac SPECT. In CDI-Net, paired projection- and image-domain networks are end-to-end connected to fuse the emission and anatomical information across domains and iterations. Adaptive Weight Recalibrators (AWR) adjust the multi-channel input features to enhance prediction accuracy. Our experiments using clinical data showed that CDI-Net produced more accurate $\mu$-maps, projections, and reconstructions compared to existing approaches that addressed each task separately. Ablation studies demonstrated the significance of cross-domain and cross-iteration connections, as well as AWR, in improving the reconstruction performance.

Abstract: Myocardial perfusion imaging (MPI) by single-photon emission computed tomography (SPECT) is widely applied for the diagnosis of cardiovascular diseases. Reducing the dose of the injected tracer is essential for lowering the patient's radiation exposure, but it will lead to increased image noise. Additionally, the latest dedicated cardiac SPECT scanners typically acquire projections in fewer angles using fewer detectors to reduce hardware expenses, potentially resulting in lower reconstruction accuracy. To overcome these challenges, we propose a dual-domain iterative network for end-to-end joint denoising and reconstruction from low-dose and few-angle projections of cardiac SPECT. The image-domain network provides a prior estimate for the projection-domain networks. The projection-domain primary and auxiliary modules are interconnected for progressive denoising and few-angle reconstruction. Adaptive Data Consistency (ADC) modules improve prediction accuracy by efficiently fusing the outputs of the primary and auxiliary modules. Experiments using clinical MPI data show that our proposed method outperforms existing image-, projection-, and dual-domain techniques, producing more accurate projections and reconstructions. Ablation studies confirm the significance of the image-domain prior estimate and ADC modules in enhancing network performance.

Abstract: Semi-supervised learning has been well developed to help reduce the cost of manual labelling by exploiting a large quantity of unlabelled data. Especially in the application of land cover classification, pixel-level manual labelling in large-scale imagery is labour-intensive and expensive. However, the existing semi-supervised learning methods pay limited attention to the quality of pseudo-labels whilst supervising the network. That is, nevertheless, one of the critical factors determining network performance. In order to fill this gap, we develop a confidence-guided semi-supervised learning (CGSSL) approach to make use of high-confidence pseudo labels and reduce the negative effect of low-confidence ones on training the land cover classification network. Meanwhile, the proposed semi-supervised learning approach uses multiple network architectures to increase pseudo-label diversity. The proposed semi-supervised learning approach significantly improves the performance of land cover classification compared to the classical semi-supervised learning methods in computer vision and even outperforms fully supervised learning with a complete set of labelled imagery of the benchmark Potsdam land cover data set.

Abstract: Inspired by the superior language abilities of large language models (LLM), large vision-language models (LVLM) have been recently explored by integrating powerful LLMs for improving the performance on complex multimodal tasks. Despite the promising progress on LVLMs, we find that LVLMs suffer from the hallucination problem, i.e. they tend to generate objects that are inconsistent with the target images in the descriptions. To investigate it, this work presents the first systematic study on object hallucination of LVLMs. We conduct the evaluation experiments on several representative LVLMs, and show that they mostly suffer from severe object hallucination issue. We further discuss that the visual instructions may influence the hallucination, and find that: objects that frequently occur in the visual instructions or co-occur with the image objects, are obviously prone to be hallucinated by LVLMs. Besides, we find that existing evaluation methods might be affected by the input instructions and generation styles of LVLMs. Thus, we further design an improved evaluation method for object hallucination by proposing a polling-based query method called \emph{POPE}. Experiment results demonstrate that our POPE can evaluate the object hallucination in a more stable and flexible way. Our codes and data are publicly available at https://github.com/RUCAIBox/POPE.

Abstract: In this paper, we focus on the problem of Medical Visual Question Answering (MedVQA), which is crucial in efficiently interpreting medical images with vital clinic-relevant information. Firstly, we reframe the problem of MedVQA as a generation task that naturally follows the human-machine interaction, we propose a generative-based model for medical visual understanding by aligning visual information from a pre-trained vision encoder with a large language model. Secondly, we establish a scalable pipeline to construct a large-scale medical visual question-answering dataset, named PMC-VQA, which contains 227k VQA pairs of 149k images that cover various modalities or diseases. Thirdly, we pre-train our proposed model on PMC-VQA and then fine-tune it on multiple public benchmarks, e.g., VQA-RAD and SLAKE, outperforming existing work by a large margin. Additionally, we propose a test set that has undergone manual verification, which is significantly more challenging, even the best models struggle to solve.

Abstract: Mimicking realistic dynamics in 3D garment animations is a challenging task due to the complex nature of multi-layered garments and the variety of outer forces involved. Existing approaches mostly focus on single-layered garments driven by only human bodies and struggle to handle general scenarios. In this paper, we propose a novel data-driven method, called LayersNet, to model garment-level animations as particle-wise interactions in a micro physics system. We improve simulation efficiency by representing garments as patch-level particles in a two-level structural hierarchy. Moreover, we introduce a novel Rotation Equivalent Transformation that leverages the rotation invariance and additivity of physics systems to better model outer forces. To verify the effectiveness of our approach and bridge the gap between experimental environments and real-world scenarios, we introduce a new challenging dataset, D-LAYERS, containing 700K frames of dynamics of 4,900 different combinations of multi-layered garments driven by both human bodies and randomly sampled wind. Our experiments show that LayersNet achieves superior performance both quantitatively and qualitatively. We will make the dataset and code publicly available at https://mmlab-ntu.github.io/project/layersnet/index.html .

Abstract: Generalized Category Discovery (GCD) requires a model to both classify known categories and cluster unknown categories in unlabeled data. Prior methods leveraged self-supervised pre-training combined with supervised fine-tuning on the labeled data, followed by simple clustering methods. In this paper, we posit that such methods are still prone to poor performance on out-of-distribution categories, and do not leverage a key ingredient: Semantic relationships between object categories. We therefore propose to leverage multi-modal (vision and language) models, in two complementary ways. First, we establish a strong baseline by replacing uni-modal features with CLIP, inspired by its zero-shot performance. Second, we propose a novel retrieval-based mechanism that leverages CLIP's aligned vision-language representations by mining text descriptions from a text corpus for the labeled and unlabeled set. We specifically use the alignment between CLIP's visual encoding of the image and textual encoding of the corpus to retrieve top-k relevant pieces of text and incorporate their embeddings to perform joint image+text semi-supervised clustering. We perform rigorous experimentation and ablations (including on where to retrieve from, how much to retrieve, and how to combine information), and validate our results on several datasets including out-of-distribution domains, demonstrating state-of-art results.

Abstract: Scene flow estimation is the task of describing the 3D motion field between temporally successive point clouds. State-of-the-art methods use strong priors and test-time optimization techniques, but require on the order of tens of seconds for large-scale point clouds, making them unusable as computer vision primitives for real-time applications such as open world object detection. Feed forward methods are considerably faster, running on the order of tens to hundreds of milliseconds for large-scale point clouds, but require expensive human supervision. To address both limitations, we propose Scene Flow via Distillation, a simple distillation framework that uses a label-free optimization method to produce pseudo-labels to supervise a feed forward model. Our instantiation of this framework, ZeroFlow, produces scene flow estimates in real-time on large-scale point clouds at quality competitive with state-of-the-art methods while using zero human labels. Notably, at test-time ZeroFlow is over 1000$\times$ faster than label-free state-of-the-art optimization-based methods on large-scale point clouds and over 1000$\times$ cheaper to train on unlabeled data compared to the cost of human annotation of that data. To facilitate research reuse, we release our code, trained model weights, and high quality pseudo-labels for the Argoverse 2 and Waymo Open datasets.

Abstract: Modern autonomous driving systems are typically divided into three main tasks: perception, prediction, and planning. The planning task involves predicting the trajectory of the ego vehicle based on inputs from both internal intention and the external environment, and manipulating the vehicle accordingly. Most existing works evaluate their performance on the nuScenes dataset using the L2 error and collision rate between the predicted trajectories and the ground truth. In this paper, we reevaluate these existing evaluation metrics and explore whether they accurately measure the superiority of different methods. Specifically, we design an MLP-based method that takes raw sensor data (e.g., past trajectory, velocity, etc.) as input and directly outputs the future trajectory of the ego vehicle, without using any perception or prediction information such as camera images or LiDAR. Surprisingly, such a simple method achieves state-of-the-art end-to-end planning performance on the nuScenes dataset, reducing the average L2 error by about 30%. We further conduct in-depth analysis and provide new insights into the factors that are critical for the success of the planning task on nuScenes dataset. Our observation also indicates that we need to rethink the current open-loop evaluation scheme of end-to-end autonomous driving in nuScenes. Codes are available at https://github.com/E2E-AD/AD-MLP.

Abstract: Diffusion models excel at text-to-image generation, especially in subject-driven generation for personalized images. However, existing methods are inefficient due to the subject-specific fine-tuning, which is computationally intensive and hampers efficient deployment. Moreover, existing methods struggle with multi-subject generation as they often blend features among subjects. We present FastComposer which enables efficient, personalized, multi-subject text-to-image generation without fine-tuning. FastComposer uses subject embeddings extracted by an image encoder to augment the generic text conditioning in diffusion models, enabling personalized image generation based on subject images and textual instructions with only forward passes. To address the identity blending problem in the multi-subject generation, FastComposer proposes cross-attention localization supervision during training, enforcing the attention of reference subjects localized to the correct regions in the target images. Naively conditioning on subject embeddings results in subject overfitting. FastComposer proposes delayed subject conditioning in the denoising step to maintain both identity and editability in subject-driven image generation. FastComposer generates images of multiple unseen individuals with different styles, actions, and contexts. It achieves 300$\times$-2500$\times$ speedup compared to fine-tuning-based methods and requires zero extra storage for new subjects. FastComposer paves the way for efficient, personalized, and high-quality multi-subject image creation. Code, model, and dataset are available at https://github.com/mit-han-lab/fastcomposer.