Computer Vision and Pattern Recognition (cs.CV)

Abstract: Twenty-five hundred years ago, the paperwork of the Achaemenid Empire was recorded on clay tablets. In 1933, archaeologists from the University of Chicago's Oriental Institute (OI) found tens of thousands of these tablets and fragments during the excavation of Persepolis. Many of these tablets have been painstakingly photographed and annotated by expert cuneiformists, and now provide a rich dataset consisting of over 5,000 annotated tablet images and 100,000 cuneiform sign bounding boxes. We leverage this dataset to develop DeepScribe, a modular computer vision pipeline capable of localizing cuneiform signs and providing suggestions for the identity of each sign. We investigate the difficulty of learning subtasks relevant to cuneiform tablet transcription on ground-truth data, finding that a RetinaNet object detector can achieve a localization mAP of 0.78 and a ResNet classifier can achieve a top-5 sign classification accuracy of 0.89. The end-to-end pipeline achieves a top-5 classification accuracy of 0.80. As part of the classification module, DeepScribe groups cuneiform signs into morphological clusters. We consider how this automatic clustering approach differs from the organization of standard, printed sign lists and what we may learn from it. These components, trained individually, are sufficient to produce a system that can analyze photos of cuneiform tablets from the Achaemenid period and provide useful transliteration suggestions to researchers. We evaluate the model's end-to-end performance on locating and classifying signs, providing a roadmap to a linguistically-aware transliteration system, then consider the model's potential utility when applied to other periods of cuneiform writing.

Abstract: The tremendous recent advances in generative artificial intelligence techniques have led to significant successes and promise in a wide range of different applications ranging from conversational agents and textual content generation to voice and visual synthesis. Amid the rise in generative AI and its increasing widespread adoption, there has been significant growing concern over the use of generative AI for malicious purposes. In the realm of visual content synthesis using generative AI, key areas of significant concern has been image forgery (e.g., generation of images containing or derived from copyright content), and data poisoning (i.e., generation of adversarially contaminated images). Motivated to address these key concerns to encourage responsible generative AI, we introduce the DeepfakeArt Challenge, a large-scale challenge benchmark dataset designed specifically to aid in the building of machine learning algorithms for generative AI art forgery and data poisoning detection. Comprising of over 32,000 records across a variety of generative forgery and data poisoning techniques, each entry consists of a pair of images that are either forgeries / adversarially contaminated or not. Each of the generated images in the DeepfakeArt Challenge benchmark dataset has been quality checked in a comprehensive manner. The DeepfakeArt Challenge is a core part of GenAI4Good, a global open source initiative for accelerating machine learning for promoting responsible creation and deployment of generative AI for good.

Abstract: We present a novel vision-language prompt learning approach for few-shot out-of-distribution (OOD) detection. Few-shot OOD detection aims to detect OOD images from classes that are unseen during training using only a few labeled in-distribution (ID) images. While prompt learning methods such as CoOp have shown effectiveness and efficiency in few-shot ID classification, they still face limitations in OOD detection due to the potential presence of ID-irrelevant information in text embeddings. To address this issue, we introduce a new approach called \textbf{Lo}cal regularized \textbf{Co}ntext \textbf{Op}timization (LoCoOp), which performs OOD regularization that utilizes the portions of CLIP local features as OOD features during training. CLIP's local features have a lot of ID-irrelevant nuisances (e.g., backgrounds), and by learning to push them away from the ID class text embeddings, we can remove the nuisances in the ID class text embeddings and enhance the separation between ID and OOD. Experiments on the large-scale ImageNet OOD detection benchmarks demonstrate the superiority of our LoCoOp over zero-shot, fully supervised detection methods and prompt learning methods. Notably, even in a one-shot setting -- just one label per class, LoCoOp outperforms existing zero-shot and fully supervised detection methods. The code will be available via \url{https://github.com/AtsuMiyai/LoCoOp}.

Abstract: Monolithic neural networks that make use of a single set of weights to learn useful representations for downstream tasks explicitly dismiss the compositional nature of data generation processes. This characteristic exists in data where every instance can be regarded as the combination of an identity concept, such as the shape of an object, combined with modifying concepts, such as orientation, color, and size. The dismissal of compositionality is especially detrimental in robotics, where state estimation relies heavily on the compositional nature of physical mechanisms (e.g., rotations and transformations) to model interactions. To accommodate this data characteristic, modular networks have been proposed. However, a lack of structure in each module's role, and modular network-specific issues such as module collapse have restricted their usability. We propose a modular network architecture that accommodates the mentioned decompositional concept by proposing a unique structure that splits the modules into predetermined roles. Additionally, we provide regularizations that improve the resiliency of the modular network to the problem of module collapse while improving the decomposition accuracy of the model.

Abstract: Manual medical image segmentation is subjective and suffers from annotator-related bias, which can be mimicked or amplified by deep learning methods. Recently, researchers have suggested that such bias is the combination of the annotator preference and stochastic error, which are modeled by convolution blocks located after decoder and pixel-wise independent Gaussian distribution, respectively. It is unlikely that convolution blocks can effectively model the varying degrees of preference at the full resolution level. Additionally, the independent pixel-wise Gaussian distribution disregards pixel correlations, leading to a discontinuous boundary. This paper proposes a Transformer-based Annotation Bias-aware (TAB) medical image segmentation model, which tackles the annotator-related bias via modeling annotator preference and stochastic errors. TAB employs the Transformer with learnable queries to extract the different preference-focused features. This enables TAB to produce segmentation with various preferences simultaneously using a single segmentation head. Moreover, TAB takes the multivariant normal distribution assumption that models pixel correlations, and learns the annotation distribution to disentangle the stochastic error. We evaluated our TAB on an OD/OC segmentation benchmark annotated by six annotators. Our results suggest that TAB outperforms existing medical image segmentation models which take into account the annotator-related bias.

Abstract: Enhancing images in low-light scenes is a challenging but widely concerned task in the computer vision. The mainstream learning-based methods mainly acquire the enhanced model by learning the data distribution from the specific scenes, causing poor adaptability (even failure) when meeting real-world scenarios that have never been encountered before. The main obstacle lies in the modeling conundrum from distribution discrepancy across different scenes. To remedy this, we first explore relationships between diverse low-light scenes based on statistical analysis, i.e., the network parameters of the encoder trained in different data distributions are close. We introduce the bilevel paradigm to model the above latent correspondence from the perspective of hyperparameter optimization. A bilevel learning framework is constructed to endow the scene-irrelevant generality of the encoder towards diverse scenes (i.e., freezing the encoder in the adaptation and testing phases). Further, we define a reinforced bilevel learning framework to provide a meta-initialization for scene-specific decoder to further ameliorate visual quality. Moreover, to improve the practicability, we establish a Retinex-induced architecture with adaptive denoising and apply our built learning framework to acquire its parameters by using two training losses including supervised and unsupervised forms. Extensive experimental evaluations on multiple datasets verify our adaptability and competitive performance against existing state-of-the-art works. The code and datasets will be available at https://github.com/vis-opt-group/BL.

Abstract: We present a generalizable novel view synthesis method where it is possible to modify the visual appearance of rendered views to match a target weather or lighting condition. Our method is based on a generalizable transformer architecture, trained on synthetically generated scenes under different appearance conditions. This allows for rendering novel views in a consistent manner of 3D scenes that were not included in the training set, along with the ability to (i) modify their appearance to match the target condition and (ii) smoothly interpolate between different conditions. Experiments on both real and synthetic scenes are provided including both qualitative and quantitative evaluations. Please refer to our project page for video results: https://ava-nvs.github.io/

Abstract: For any digital application with document images such as retrieval, the classification of document images becomes an essential stage. Conventionally for the purpose, the full versions of the documents, that is the uncompressed document images make the input dataset, which poses a threat due to the big volume required to accommodate the full versions of the documents. Therefore, it would be novel, if the same classification task could be accomplished directly (with some partial decompression) with the compressed representation of documents in order to make the whole process computationally more efficient. In this research work, a novel deep learning model, DWT CompCNN is proposed for classification of documents that are compressed using High Throughput JPEG 2000 (HTJ2K) algorithm. The proposed DWT-CompCNN comprises of five convolutional layers with filter sizes of 16, 32, 64, 128, and 256 consecutively for each increasing layer to improve learning from the wavelet coefficients extracted from the compressed images. Experiments are performed on two benchmark datasets- Tobacco-3482 and RVL-CDIP, which demonstrate that the proposed model is time and space efficient, and also achieves a better classification accuracy in compressed domain.

Abstract: Recent advances in score-based generative models have led to a huge spike in the development of downstream applications using generative models ranging from data augmentation over image and video generation to anomaly detection. Despite publicly available trained models, their potential to be used for privacy preserving data sharing has not been fully explored yet. Training diffusion models on private data and disseminating the models and weights rather than the raw dataset paves the way for innovative large-scale data-sharing strategies, particularly in healthcare, where safeguarding patients' personal health information is paramount. However, publishing such models without individual consent of, e.g., the patients from whom the data was acquired, necessitates guarantees that identifiable training samples will never be reproduced, thus protecting personal health data and satisfying the requirements of policymakers and regulatory bodies. This paper introduces a method for estimating the upper bound of the probability of reproducing identifiable training images during the sampling process. This is achieved by designing an adversarial approach that searches for anatomic fingerprints, such as medical devices or dermal art, which could potentially be employed to re-identify training images. Our method harnesses the learned score-based model to estimate the probability of the entire subspace of the score function that may be utilized for one-to-one reproduction of training samples. To validate our estimates, we generate anomalies containing a fingerprint and investigate whether generated samples from trained generative models can be uniquely mapped to the original training samples. Overall our results show that privacy-breaching images are reproduced at sampling time if the models were trained without care.

Abstract: Summarizing a video requires a diverse understanding of the video, ranging from recognizing scenes to evaluating how much each frame is essential enough to be selected as a summary. Self-supervised learning (SSL) is acknowledged for its robustness and flexibility to multiple downstream tasks, but the video SSL has not shown its value for dense understanding tasks like video summarization. We claim an unsupervised autoencoder with sufficient self-supervised learning does not need any extra downstream architecture design or fine-tuning weights to be utilized as a video summarization model. The proposed method to evaluate the importance score of each frame takes advantage of the reconstruction score of the autoencoder's decoder. We evaluate the method in major unsupervised video summarization benchmarks to show its effectiveness under various experimental settings.

Abstract: Despite impressive empirical advances of SSL in solving various tasks, the problem of understanding and characterizing SSL representations learned from input data remains relatively under-explored. We provide a comparative analysis of how the representations produced by SSL models differ when masking parts of the input. Specifically, we considered state-of-the-art SSL pretrained models, such as DINOv2, MAE, and SwaV, and analyzed changes at the representation levels across 4 Image Classification datasets. First, we generate variations of the datasets by applying foreground and background segmentation. Then, we conduct statistical analysis using Canonical Correlation Analysis (CCA) and Centered Kernel Alignment (CKA) to evaluate the robustness of the representations learned in SSL models. Empirically, we show that not all models lead to representations that separate foreground, background, and complete images. Furthermore, we test different masking strategies by occluding the center regions of the images to address cases where foreground and background are difficult. For example, the DTD dataset that focuses on texture rather specific objects.

Abstract: Learning signed distance functions (SDFs) from 3D point clouds is an important task in 3D computer vision. However, without ground truth signed distances, point normals or clean point clouds, current methods still struggle from learning SDFs from noisy point clouds. To overcome this challenge, we propose to learn SDFs via a noise to noise mapping, which does not require any clean point cloud or ground truth supervision for training. Our novelty lies in the noise to noise mapping which can infer a highly accurate SDF of a single object or scene from its multiple or even single noisy point cloud observations. Our novel learning manner is supported by modern Lidar systems which capture multiple noisy observations per second. We achieve this by a novel loss which enables statistical reasoning on point clouds and maintains geometric consistency although point clouds are irregular, unordered and have no point correspondence among noisy observations. Our evaluation under the widely used benchmarks demonstrates our superiority over the state-of-the-art methods in surface reconstruction, point cloud denoising and upsampling. Our code, data, and pre-trained models are available at https://github.com/mabaorui/Noise2NoiseMapping/

Abstract: This paper presents a novel approach for generating 3D talking heads from raw audio inputs. Our method grounds on the idea that speech related movements can be comprehensively and efficiently described by the motion of a few control points located on the movable parts of the face, i.e., landmarks. The underlying musculoskeletal structure then allows us to learn how their motion influences the geometrical deformations of the whole face. The proposed method employs two distinct models to this aim: the first one learns to generate the motion of a sparse set of landmarks from the given audio. The second model expands such landmarks motion to a dense motion field, which is utilized to animate a given 3D mesh in neutral state. Additionally, we introduce a novel loss function, named Cosine Loss, which minimizes the angle between the generated motion vectors and the ground truth ones. Using landmarks in 3D talking head generation offers various advantages such as consistency, reliability, and obviating the need for manual-annotation. Our approach is designed to be identity-agnostic, enabling high-quality facial animations for any users without additional data or training.

Abstract: Network optimization is a crucial step in the field of deep learning, as it directly affects the performance of models in various domains such as computer vision. Despite the numerous optimizers that have been developed over the years, the current methods are still limited in their ability to accurately and quickly identify gradient trends, which can lead to sub-optimal network performance. In this paper, we propose a novel deep optimizer called Fast-Adaptive Moment Estimation (FAME), which for the first time estimates gradient moments using a Triple Exponential Moving Average (TEMA). Incorporating TEMA into the optimization process provides richer and more accurate information on data changes and trends, as compared to the standard Exponential Moving Average used in essentially all current leading adaptive optimization methods. Our proposed FAME optimizer has been extensively validated through a wide range of benchmarks, including CIFAR-10, CIFAR-100, PASCAL-VOC, MS-COCO, and Cityscapes, using 14 different learning architectures, six optimizers, and various vision tasks, including detection, classification and semantic understanding. The results demonstrate that our FAME optimizer outperforms other leading optimizers in terms of both robustness and accuracy.

Abstract: LiDAR and Radar are two complementary sensing approaches in that LiDAR specializes in capturing an object's 3D shape while Radar provides longer detection ranges as well as velocity hints. Though seemingly natural, how to efficiently combine them for improved feature representation is still unclear. The main challenge arises from that Radar data are extremely sparse and lack height information. Therefore, directly integrating Radar features into LiDAR-centric detection networks is not optimal. In this work, we introduce a bi-directional LiDAR-Radar fusion framework, termed Bi-LRFusion, to tackle the challenges and improve 3D detection for dynamic objects. Technically, Bi-LRFusion involves two steps: first, it enriches Radar's local features by learning important details from the LiDAR branch to alleviate the problems caused by the absence of height information and extreme sparsity; second, it combines LiDAR features with the enhanced Radar features in a unified bird's-eye-view representation. We conduct extensive experiments on nuScenes and ORR datasets, and show that our Bi-LRFusion achieves state-of-the-art performance for detecting dynamic objects. Notably, Radar data in these two datasets have different formats, which demonstrates the generalizability of our method. Codes are available at https://github.com/JessieW0806/BiLRFusion.

Abstract: Nowadays, deploying a robust face recognition product becomes easy with the development of face recognition techniques for decades. Not only profile image verification but also the state-of-the-art method can handle the in-the-wild image almost perfectly. However, the concern of privacy issues raise rapidly since mainstream research results are powered by tons of web-crawled data, which faces the privacy invasion issue. The community tries to escape this predicament completely by training the face recognition model with synthetic data but faces severe domain gap issues, which still need to access real images and identity labels to fine-tune the model. In this paper, we propose SASMU, a simple, novel, and effective method for face recognition using a synthetic dataset. Our proposed method consists of spatial data augmentation (SA) and spectrum mixup (SMU). We first analyze the existing synthetic datasets for developing a face recognition system. Then, we reveal that heavy data augmentation is helpful for boosting performance when using synthetic data. By analyzing the previous frequency mixup studies, we proposed a novel method for domain generalization. Extensive experimental results have demonstrated the effectiveness of SASMU, achieving state-of-the-art performance on several common benchmarks, such as LFW, AgeDB-30, CA-LFW, CFP-FP, and CP-LFW.

Abstract: Cutting out an object and estimating its opacity mask, known as image matting, is a key task in image and video editing. Due to the highly ill-posed issue, additional inputs, typically user-defined trimaps or scribbles, are usually needed to reduce the uncertainty. Although effective, it is either time consuming or only suitable for experienced users who know where to place the strokes. In this work, we propose a decomposed-uncertainty-guided matting (dugMatting) algorithm, which explores the explicitly decomposed uncertainties to efficiently and effectively improve the results. Basing on the characteristic of these uncertainties, the epistemic uncertainty is reduced in the process of guiding interaction (which introduces prior knowledge), while the aleatoric uncertainty is reduced in modeling data distribution (which introduces statistics for both data and possible noise). The proposed matting framework relieves the requirement for users to determine the interaction areas by using simple and efficient labeling. Extensively quantitative and qualitative results validate that the proposed method significantly improves the original matting algorithms in terms of both efficiency and efficacy.

Abstract: This paper presents PolyDiffuse, a novel structured reconstruction algorithm that transforms visual sensor data into polygonal shapes with Diffusion Models (DM), an emerging machinery amid exploding generative AI, while formulating reconstruction as a generation process conditioned on sensor data. The task of structured reconstruction poses two fundamental challenges to DM: 1) A structured geometry is a ``set'' (e.g., a set of polygons for a floorplan geometry), where a sample of $N$ elements has $N!$ different but equivalent representations, making the denoising highly ambiguous; and 2) A ``reconstruction'' task has a single solution, where an initial noise needs to be chosen carefully, while any initial noise works for a generation task. Our technical contribution is the introduction of a Guided Set Diffusion Model where 1) the forward diffusion process learns guidance networks to control noise injection so that one representation of a sample remains distinct from its other permutation variants, thus resolving denoising ambiguity; and 2) the reverse denoising process reconstructs polygonal shapes, initialized and directed by the guidance networks, as a conditional generation process subject to the sensor data. We have evaluated our approach for reconstructing two types of polygonal shapes: floorplan as a set of polygons and HD map for autonomous cars as a set of polylines. Through extensive experiments on standard benchmarks, we demonstrate that PolyDiffuse significantly advances the current state of the art and enables broader practical applications.

Abstract: Reference-based super-resolution (RefSR) has gained considerable success in the field of super-resolution with the addition of high-resolution reference images to reconstruct low-resolution (LR) inputs with more high-frequency details, thereby overcoming some limitations of single image super-resolution (SISR). Previous research in the field of RefSR has mostly focused on two crucial aspects. The first is accurate correspondence matching between the LR and the reference (Ref) image. The second is the effective transfer and aggregation of similar texture information from the Ref images. Nonetheless, an important detail of perceptual loss and adversarial loss has been underestimated, which has a certain adverse effect on texture transfer and reconstruction. In this study, we propose a feature reuse framework that guides the step-by-step texture reconstruction process through different stages, reducing the negative impacts of perceptual and adversarial loss. The feature reuse framework can be used for any RefSR model, and several RefSR approaches have improved their performance after being retrained using our framework. Additionally, we introduce a single image feature embedding module and a texture-adaptive aggregation module. The single image feature embedding module assists in reconstructing the features of the LR inputs itself and effectively lowers the possibility of including irrelevant textures. The texture-adaptive aggregation module dynamically perceives and aggregates texture information between the LR inputs and the Ref images using dynamic filters. This enhances the utilization of the reference texture while reducing reference misuse. The source code is available at https://github.com/Yi-Yang355/FRFSR.

Abstract: In this paper, we introduce a novel Synchronized Class Token Fusion (SCT Fusion) architecture in the framework of multi-modal multi-label classification (MLC) of remote sensing (RS) images. The proposed architecture leverages modality-specific attention-based transformer encoders to process varying input modalities, while exchanging information across modalities by synchronizing the special class tokens after each transformer encoder block. The synchronization involves fusing the class tokens with a trainable fusion transformation, resulting in a synchronized class token that contains information from all modalities. As the fusion transformation is trainable, it allows to reach an accurate representation of the shared features among different modalities. Experimental results show the effectiveness of the proposed architecture over single-modality architectures and an early fusion multi-modal architecture when evaluated on a multi-modal MLC dataset. The code of the proposed architecture is publicly available at https://git.tu-berlin.de/rsim/sct-fusion.

Abstract: Due to the over-parameterization of neural networks, many model compression methods based on pruning and quantization have emerged. They are remarkable in reducing the size, parameter number, and computational complexity of the model. However, most of the models compressed by such methods need the support of special hardware and software, which increases the deployment cost. Moreover, these methods are mainly used in classification tasks, and rarely directly used in detection tasks. To address these issues, for the object detection network we introduce a three-stage model compression method: dynamic sparse training, group channel pruning, and spatial attention distilling. Firstly, to select out the unimportant channels in the network and maintain a good balance between sparsity and accuracy, we put forward a dynamic sparse training method, which introduces a variable sparse rate, and the sparse rate will change with the training process of the network. Secondly, to reduce the effect of pruning on network accuracy, we propose a novel pruning method called group channel pruning. In particular, we divide the network into multiple groups according to the scales of the feature layer and the similarity of module structure in the network, and then we use different pruning thresholds to prune the channels in each group. Finally, to recover the accuracy of the pruned network, we use an improved knowledge distillation method for the pruned network. Especially, we extract spatial attention information from the feature maps of specific scales in each group as knowledge for distillation. In the experiments, we use YOLOv4 as the object detection network and PASCAL VOC as the training dataset. Our method reduces the parameters of the model by 64.7 % and the calculation by 34.9%.

Abstract: Achieving an immersive experience enabling users to explore virtual environments with six degrees of freedom (6DoF) is essential for various applications such as virtual reality (VR). Wide-baseline panoramas are commonly used in these applications to reduce network bandwidth and storage requirements. However, synthesizing novel views from these panoramas remains a key challenge. Although existing neural radiance field methods can produce photorealistic views under narrow-baseline and dense image captures, they tend to overfit the training views when dealing with \emph{wide-baseline} panoramas due to the difficulty in learning accurate geometry from sparse $360^{\circ}$ views. To address this problem, we propose PanoGRF, Generalizable Spherical Radiance Fields for Wide-baseline Panoramas, which construct spherical radiance fields incorporating $360^{\circ}$ scene priors. Unlike generalizable radiance fields trained on perspective images, PanoGRF avoids the information loss from panorama-to-perspective conversion and directly aggregates geometry and appearance features of 3D sample points from each panoramic view based on spherical projection. Moreover, as some regions of the panorama are only visible from one view while invisible from others under wide baseline settings, PanoGRF incorporates $360^{\circ}$ monocular depth priors into spherical depth estimation to improve the geometry features. Experimental results on multiple panoramic datasets demonstrate that PanoGRF significantly outperforms state-of-the-art generalizable view synthesis methods for wide-baseline panoramas (e.g., OmniSyn) and perspective images (e.g., IBRNet, NeuRay).

Abstract: The recent Segment Anything Model (SAM) represents a big leap in scaling up segmentation models, allowing for powerful zero-shot capabilities and flexible prompting. Despite being trained with 1.1 billion masks, SAM's mask prediction quality falls short in many cases, particularly when dealing with objects that have intricate structures. We propose HQ-SAM, equipping SAM with the ability to accurately segment any object, while maintaining SAM's original promptable design, efficiency, and zero-shot generalizability. Our careful design reuses and preserves the pre-trained model weights of SAM, while only introducing minimal additional parameters and computation. We design a learnable High-Quality Output Token, which is injected into SAM's mask decoder and is responsible for predicting the high-quality mask. Instead of only applying it on mask-decoder features, we first fuse them with early and final ViT features for improved mask details. To train our introduced learnable parameters, we compose a dataset of 44K fine-grained masks from several sources. HQ-SAM is only trained on the introduced detaset of 44k masks, which takes only 4 hours on 8 GPUs. We show the efficacy of HQ-SAM in a suite of 9 diverse segmentation datasets across different downstream tasks, where 7 out of them are evaluated in a zero-shot transfer protocol. Our code and models will be released at https://github.com/SysCV/SAM-HQ.

Abstract: Biomedical image classification requires capturing of bio-informatics based on specific feature distribution. In most of such applications, there are mainly challenges due to limited availability of samples for diseased cases and imbalanced nature of dataset. This article presents the novel framework of multi-head self-attention for vision transformer (ViT) which makes capable of capturing the specific image features for classification and analysis. The proposed method uses the concept of residual connection for accumulating the best attention output in each block of multi-head attention. The proposed framework has been evaluated on two small datasets: (i) blood cell classification dataset and (ii) brain tumor detection using brain MRI images. The results show the significant improvement over traditional ViT and other convolution based state-of-the-art classification models.

Abstract: Camera pose estimation for two-view geometry traditionally relies on RANSAC. Normally, a multitude of image correspondences leads to a pool of proposed hypotheses, which are then scored to find a winning model. The inlier count is generally regarded as a reliable indicator of "consensus". We examine this scoring heuristic, and find that it favors disappointing models under certain circumstances. As a remedy, we propose the Fundamental Scoring Network (FSNet), which infers a score for a pair of overlapping images and any proposed fundamental matrix. It does not rely on sparse correspondences, but rather embodies a two-view geometry model through an epipolar attention mechanism that predicts the pose error of the two images. FSNet can be incorporated into traditional RANSAC loops. We evaluate FSNet on fundamental and essential matrix estimation on indoor and outdoor datasets, and establish that FSNet can successfully identify good poses for pairs of images with few or unreliable correspondences. Besides, we show that naively combining FSNet with MAGSAC++ scoring approach achieves state of the art results.

Abstract: Domain adaptive semantic segmentation enables robust pixel-wise understanding in real-world driving scenes. Source-free domain adaptation, as a more practical technique, addresses the concerns of data privacy and storage limitations in typical unsupervised domain adaptation methods. It utilizes a well-trained source model and unlabeled target data to achieve adaptation in the target domain. However, in the absence of source data and target labels, current solutions cannot sufficiently reduce the impact of domain shift and fully leverage the information from the target data. In this paper, we propose an end-to-end source-free domain adaptation semantic segmentation method via Importance-Aware and Prototype-Contrast (IAPC) learning. The proposed IAPC framework effectively extracts domain-invariant knowledge from the well-trained source model and learns domain-specific knowledge from the unlabeled target domain. Specifically, considering the problem of domain shift in the prediction of the target domain by the source model, we put forward an importance-aware mechanism for the biased target prediction probability distribution to extract domain-invariant knowledge from the source model. We further introduce a prototype-contrast strategy, which includes a prototype-symmetric cross-entropy loss and a prototype-enhanced cross-entropy loss, to learn target intra-domain knowledge without relying on labels. A comprehensive variety of experiments on two domain adaptive semantic segmentation benchmarks demonstrates that the proposed end-to-end IAPC solution outperforms existing state-of-the-art methods. Code will be made publicly available at https://github.com/yihong-97/Source-free_IAPC.

Abstract: Recent advances in self-supervised representation learning have enabled more efficient and robust model performance without relying on extensive labeled data. However, most works are still focused on images, with few working on videos and even fewer on multi-view videos, where more powerful inductive biases can be leveraged for self-supervision. In this work, we propose a novel method for representation learning of multi-view videos, where we explicitly model the representation space to maintain Homography Equivariance (HomE). Our method learns an implicit mapping between different views, culminating in a representation space that maintains the homography relationship between neighboring views. We evaluate our HomE representation via action recognition and pedestrian intent prediction as downstream tasks. On action classification, our method obtains 96.4% 3-fold accuracy on the UCF101 dataset, better than most state-of-the-art self-supervised learning methods. Similarly, on the STIP dataset, we outperform the state-of-the-art by 6% for pedestrian intent prediction one second into the future while also obtaining an accuracy of 91.2% for pedestrian action (cross vs. not-cross) classification. Code is available at https://github.com/anirudhs123/HomE.

Abstract: Digitalization of existing buildings and the creation of 3D BIM models for them has become crucial for many tasks. Of particular importance are floor plans, which contain information about building layouts and are vital for processes such as construction, maintenance or refurbishing. However, this data is not always available in digital form, especially for older buildings constructed before CAD tools were widely available, or lacks semantic information. The digitalization of such information usually requires manual work of an expert that must reconstruct the layouts by hand, which is a cumbersome and error-prone process. In this paper, we present a pipeline for reconstruction of vectorized 3D models from scanned 2D plans, aiming at increasing the efficiency of this process. The method presented achieves state-of-the-art results in the public dataset CubiCasa5k, and shows good generalization to different types of plans. Our vectorization approach is particularly effective, outperforming previous methods.

Abstract: Listeners use short interjections, so-called backchannels, to signify attention or express agreement. The automatic analysis of this behavior is of key importance for human conversation analysis and interactive conversational agents. Current state-of-the-art approaches for backchannel analysis from visual behavior make use of two types of features: features based on body pose and features based on facial behavior. At the same time, transformer neural networks have been established as an effective means to fuse input from different data sources, but they have not yet been applied to backchannel analysis. In this work, we conduct a comprehensive evaluation of multi-modal transformer architectures for automatic backchannel analysis based on pose and facial information. We address both the detection of backchannels as well as the task of estimating the agreement expressed in a backchannel. In evaluations on the MultiMediate'22 backchannel detection challenge, we reach 66.4% accuracy with a one-layer transformer architecture, outperforming the previous state of the art. With a two-layer transformer architecture, we furthermore set a new state of the art (0.0604 MSE) on the task of estimating the amount of agreement expressed in a backchannel.

Abstract: In-context learning$\unicode{x2013}$the ability to configure a model's behavior with different prompts$\unicode{x2013}$has revolutionized the field of natural language processing, alleviating the need for task-specific models and paving the way for generalist models capable of assisting with any query. Computer vision, in contrast, has largely stayed in the former regime: specialized decoders and finetuning protocols are generally required to perform dense tasks such as semantic segmentation and depth estimation. In this work we explore a simple mechanism for in-context learning of such scene understanding tasks: nearest neighbor retrieval from a prompt of annotated features. We propose a new pretraining protocol$\unicode{x2013}$leveraging attention within and across images$\unicode{x2013}$which yields representations particularly useful in this regime. The resulting Hummingbird model, suitably prompted, performs various scene understanding tasks without modification while approaching the performance of specialists that have been finetuned for each task. Moreover, Hummingbird can be configured to perform new tasks much more efficiently than finetuned models, raising the possibility of scene understanding in the interactive assistant regime.

Abstract: Fine-tuning vision-language models (VLMs) like CLIP to downstream tasks is often necessary to optimize their performance. However, a major obstacle is the limited availability of labeled data. We study the use of pseudolabels, i.e., heuristic labels for unlabeled data, to enhance CLIP via prompt tuning. Conventional pseudolabeling trains a model on labeled data and then generates labels for unlabeled data. VLMs' zero-shot capabilities enable a ``second generation'' of pseudolabeling approaches that do not require task-specific training on labeled data. By using zero-shot pseudolabels as a source of supervision, we observe that learning paradigms such as semi-supervised, transductive zero-shot, and unsupervised learning can all be seen as optimizing the same loss function. This unified view enables the development of versatile training strategies that are applicable across learning paradigms. We investigate them on image classification tasks where CLIP exhibits limitations, by varying prompt modalities, e.g., textual or visual prompts, and learning paradigms. We find that (1) unexplored prompt tuning strategies that iteratively refine pseudolabels consistently improve CLIP accuracy, by 19.5 points in semi-supervised learning, by 28.4 points in transductive zero-shot learning, and by 15.2 points in unsupervised learning, and (2) unlike conventional semi-supervised pseudolabeling, which exacerbates model biases toward classes with higher-quality pseudolabels, prompt tuning leads to a more equitable distribution of per-class accuracy. The code to reproduce the experiments is at github.com/BatsResearch/menghini-enhanceCLIPwithCLIP-code.

Abstract: Unsupervised domain adaptation (UDA) aims to bridge the gap between source and target domains in the absence of target domain labels using two main techniques: input-level alignment (such as generative modeling and stylization) and feature-level alignment (which matches the distribution of the feature maps, e.g. gradient reversal layers). Motivated from the success of generative modeling for image classification, stylization-based methods were recently proposed for regression tasks, such as pose estimation. However, use of input-level alignment via generative modeling and stylization incur additional overhead and computational complexity which limit their use in real-world DA tasks. To investigate the role of input-level alignment for DA, we ask the following question: Is generative modeling-based stylization necessary for visual domain adaptation in regression? Surprisingly, we find that input-alignment has little effect on regression tasks as compared to classification. Based on these insights, we develop a non-parametric feature-level domain alignment method -- Implicit Stylization (ImSty) -- which results in consistent improvements over SOTA regression task, without the need for computationally intensive stylization and generative modeling. Our work conducts a critical evaluation of the role of generative modeling and stylization, at a time when these are also gaining popularity for domain generalization.

Abstract: The evolution of semantic segmentation has long been dominated by learning more discriminative image representations for classifying each pixel. Despite the prominent advancements, the priors of segmentation masks themselves, e.g., geometric and semantic constraints, are still under-explored. In this paper, we propose to ameliorate the semantic segmentation quality of existing discriminative approaches with a mask prior modeled by a recently-developed denoising diffusion generative model. Beginning with a unified architecture that adapts diffusion models for mask prior modeling, we focus this work on a specific instantiation with discrete diffusion and identify a variety of key design choices for its successful application. Our exploratory analysis revealed several important findings, including: (1) a simple integration of diffusion models into semantic segmentation is not sufficient, and a poorly-designed diffusion process might lead to degradation in segmentation performance; (2) during the training, the object to which noise is added is more important than the type of noise; (3) during the inference, the strict diffusion denoising scheme may not be essential and can be relaxed to a simpler scheme that even works better. We evaluate the proposed prior modeling with several off-the-shelf segmentors, and our experimental results on ADE20K and Cityscapes demonstrate that our approach could achieve competitively quantitative performance and more appealing visual quality.

Abstract: Video colorization is a challenging task that involves inferring plausible and temporally consistent colors for grayscale frames. In this paper, we present ColorDiffuser, an adaptation of a pre-trained text-to-image latent diffusion model for video colorization. With the proposed adapter-based approach, we repropose the pre-trained text-to-image model to accept input grayscale video frames, with the optional text description, for video colorization. To enhance the temporal coherence and maintain the vividness of colorization across frames, we propose two novel techniques: the Color Propagation Attention and Alternated Sampling Strategy. Color Propagation Attention enables the model to refine its colorization decision based on a reference latent frame, while Alternated Sampling Strategy captures spatiotemporal dependencies by using the next and previous adjacent latent frames alternatively as reference during the generative diffusion sampling steps. This encourages bidirectional color information propagation between adjacent video frames, leading to improved color consistency across frames. We conduct extensive experiments on benchmark datasets, and the results demonstrate the effectiveness of our proposed framework. The evaluations show that ColorDiffuser achieves state-of-the-art performance in video colorization, surpassing existing methods in terms of color fidelity, temporal consistency, and visual quality.

Abstract: We propose DocFormerv2, a multi-modal transformer for Visual Document Understanding (VDU). The VDU domain entails understanding documents (beyond mere OCR predictions) e.g., extracting information from a form, VQA for documents and other tasks. VDU is challenging as it needs a model to make sense of multiple modalities (visual, language and spatial) to make a prediction. Our approach, termed DocFormerv2 is an encoder-decoder transformer which takes as input - vision, language and spatial features. DocFormerv2 is pre-trained with unsupervised tasks employed asymmetrically i.e., two novel document tasks on encoder and one on the auto-regressive decoder. The unsupervised tasks have been carefully designed to ensure that the pre-training encourages local-feature alignment between multiple modalities. DocFormerv2 when evaluated on nine datasets shows state-of-the-art performance over strong baselines e.g. TabFact (4.3%), InfoVQA (1.4%), FUNSD (1%). Furthermore, to show generalization capabilities, on three VQA tasks involving scene-text, Doc- Formerv2 outperforms previous comparably-sized models and even does better than much larger models (such as GIT2, PaLi and Flamingo) on some tasks. Extensive ablations show that due to its pre-training, DocFormerv2 understands multiple modalities better than prior-art in VDU.

Abstract: Observing the close relationship among panoptic, semantic and instance segmentation tasks, we propose to train a universal multi-dataset multi-task segmentation model: DaTaSeg.We use a shared representation (mask proposals with class predictions) for all tasks. To tackle task discrepancy, we adopt different merge operations and post-processing for different tasks. We also leverage weak-supervision, allowing our segmentation model to benefit from cheaper bounding box annotations. To share knowledge across datasets, we use text embeddings from the same semantic embedding space as classifiers and share all network parameters among datasets. We train DaTaSeg on ADE semantic, COCO panoptic, and Objects365 detection datasets. DaTaSeg improves performance on all datasets, especially small-scale datasets, achieving 54.0 mIoU on ADE semantic and 53.5 PQ on COCO panoptic. DaTaSeg also enables weakly-supervised knowledge transfer on ADE panoptic and Objects365 instance segmentation. Experiments show DaTaSeg scales with the number of training datasets and enables open-vocabulary segmentation through direct transfer. In addition, we annotate an Objects365 instance segmentation set of 1,000 images and will release it as a public benchmark.

Abstract: Multi-view 3D object detection is becoming popular in autonomous driving due to its high effectiveness and low cost. Most of the current state-of-the-art detectors follow the query-based bird's-eye-view (BEV) paradigm, which benefits from both BEV's strong perception power and end-to-end pipeline. Despite achieving substantial progress, existing works model objects via globally leveraging temporal and spatial information of BEV features, resulting in problems when handling the challenging complex and dynamic autonomous driving scenarios. In this paper, we proposed an Object-Centric query-BEV detector OCBEV, which can carve the temporal and spatial cues of moving targets more effectively. OCBEV comprises three designs: Object Aligned Temporal Fusion aligns the BEV feature based on ego-motion and estimated current locations of moving objects, leading to a precise instance-level feature fusion. Object Focused Multi-View Sampling samples more 3D features from an adaptive local height ranges of objects for each scene to enrich foreground information. Object Informed Query Enhancement replaces part of pre-defined decoder queries in common DETR-style decoders with positional features of objects on high-confidence locations, introducing more direct object positional priors. Extensive experimental evaluations are conducted on the challenging nuScenes dataset. Our approach achieves a state-of-the-art result, surpassing the traditional BEVFormer by 1.5 NDS points. Moreover, we have a faster convergence speed and only need half of the training iterations to get comparable performance, which further demonstrates its effectiveness.