Computer Vision and Pattern Recognition (cs.CV)

Abstract: In recent years, the development of instance segmentation has garnered significant attention in a wide range of applications. However, the training of a fully-supervised instance segmentation model requires costly both instance-level and pixel-level annotations. In contrast, weakly-supervised instance segmentation methods (i.e., with image-level class labels or point labels) struggle to satisfy the accuracy and recall requirements of practical scenarios. In this paper, we propose a novel paradigm for instance segmentation called training-free instance segmentation (TFISeg), which achieves instance segmentation results from image masks predicted using off-the-shelf semantic segmentation models. TFISeg does not require training a semantic or/and instance segmentation model and avoids the need for instance-level image annotations. Therefore, it is highly efficient. Specifically, we first obtain a semantic segmentation mask of the input image via a trained semantic segmentation model. Then, we calculate a displacement field vector for each pixel based on the segmentation mask, which can indicate representations belonging to the same class but different instances, i.e., obtaining the instance-level object information. Finally, instance segmentation results are obtained after being refined by a learnable category-agnostic object boundary branch. Extensive experimental results on two challenging datasets and representative semantic segmentation baselines (including CNNs and Transformers) demonstrate that TFISeg can achieve competitive results compared to the state-of-the-art fully-supervised instance segmentation methods without the need for additional human resources or increased computational costs. The code is available at: TFISeg

Abstract: Addressing the rising concerns of privacy and security, domain adaptation in the dark aims to adapt a black-box source trained model to an unlabeled target domain without access to any source data or source model parameters. The need for domain adaptation of black-box predictors becomes even more pronounced to protect intellectual property as deep learning based solutions are becoming increasingly commercialized. Current methods distill noisy predictions on the target data obtained from the source model to the target model, and/or separate clean/noisy target samples before adapting using traditional noisy label learning algorithms. However, these methods do not utilize the easy-to-hard learning nature of the clean/noisy data splits. Also, none of the existing methods are end-to-end, and require a separate fine-tuning stage and an initial warmup stage. In this work, we present Curriculum Adaptation for Black-Box (CABB) which provides a curriculum guided adaptation approach to gradually train the target model, first on target data with high confidence (clean) labels, and later on target data with noisy labels. CABB utilizes Jensen-Shannon divergence as a better criterion for clean-noisy sample separation, compared to the traditional criterion of cross entropy loss. Our method utilizes co-training of a dual-branch network to suppress error accumulation resulting from confirmation bias. The proposed approach is end-to-end trainable and does not require any extra finetuning stage, unlike existing methods. Empirical results on standard domain adaptation datasets show that CABB outperforms existing state-of-the-art black-box DA models and is comparable to white-box domain adaptation models.

Abstract: The prominent progress in generative models has significantly improved the reality of generated faces, bringing serious concerns to society. Since recent GAN-generated faces are in high realism, the forgery traces have become more imperceptible, increasing the forensics challenge. To combat GAN-generated faces, many countermeasures based on Convolutional Neural Networks (CNNs) have been spawned due to their strong learning ability. In this paper, we rethink this problem and explore a new approach based on forest models instead of CNNs. Specifically, we describe a simple and effective forest-based method set called {\em ForensicsForest Family} to detect GAN-generate faces. The proposed ForensicsForest family is composed of three variants, which are {\em ForensicsForest}, {\em Hybrid ForensicsForest} and {\em Divide-and-Conquer ForensicsForest} respectively. ForenscisForest is a newly proposed Multi-scale Hierarchical Cascade Forest, which takes semantic, frequency and biology features as input, hierarchically cascades different levels of features for authenticity prediction, and then employs a multi-scale ensemble scheme that can comprehensively consider different levels of information to improve the performance further. Based on ForensicsForest, we develop Hybrid ForensicsForest, an extended version that integrates the CNN layers into models, to further refine the effectiveness of augmented features. Moreover, to reduce the memory cost in training, we propose Divide-and-Conquer ForensicsForest, which can construct a forest model using only a portion of training samplings. In the training stage, we train several candidate forest models using the subsets of training samples. Then a ForensicsForest is assembled by picking the suitable components from these candidate forest models...

Abstract: Retrieving relevant multimedia content is one of the main problems in a world that is increasingly data-driven. With the proliferation of drones, high quality aerial footage is now available to a wide audience for the first time. Integrating this footage into applications can enable GPS-less geo-localisation or location correction. In this paper, we present an orientation-guided training framework for UAV-view geo-localisation. Through hierarchical localisation orientations of the UAV images are estimated in relation to the satellite imagery. We propose a lightweight prediction module for these pseudo labels which predicts the orientation between the different views based on the contrastive learned embeddings. We experimentally demonstrate that this prediction supports the training and outperforms previous approaches. The extracted pseudo-labels also enable aligned rotation of the satellite image as augmentation to further strengthen the generalisation. During inference, we no longer need this orientation module, which means that no additional computations are required. We achieve state-of-the-art results on both the University-1652 and University-160k datasets.

Abstract: We live in a vast ocean of data, and deep neural networks are no exception to this. However, this data exhibits an inherent phenomenon of imbalance. This imbalance poses a risk of deep neural networks producing biased predictions, leading to potentially severe ethical and social consequences. To address these challenges, we believe that the use of generative models is a promising approach for comprehending tasks, given the remarkable advancements demonstrated by recent diffusion models in generating high-quality images. In this work, we propose a simple yet effective baseline, SYNAuG, that utilizes synthetic data as a preliminary step before employing task-specific algorithms to address data imbalance problems. This straightforward approach yields impressive performance on datasets such as CIFAR100-LT, ImageNet100-LT, UTKFace, and Waterbird, surpassing the performance of existing task-specific methods. While we do not claim that our approach serves as a complete solution to the problem of data imbalance, we argue that supplementing the existing data with synthetic data proves to be an effective and crucial preliminary step in addressing data imbalance concerns.

Abstract: Supervised 3D Object Detection models have been displaying increasingly better performance in single-domain cases where the training data comes from the same environment and sensor as the testing data. However, in real-world scenarios data from the target domain may not be available for finetuning or for domain adaptation methods. Indeed, 3D object detection models trained on a source dataset with a specific point distribution have shown difficulties in generalizing to unseen datasets. Therefore, we decided to leverage the information available from several annotated source datasets with our Multi-Dataset Training for 3D Object Detection (MDT3D) method to increase the robustness of 3D object detection models when tested in a new environment with a different sensor configuration. To tackle the labelling gap between datasets, we used a new label mapping based on coarse labels. Furthermore, we show how we managed the mix of datasets during training and finally introduce a new cross-dataset augmentation method: cross-dataset object injection. We demonstrate that this training paradigm shows improvements for different types of 3D object detection models. The source code and additional results for this research project will be publicly available on GitHub for interested parties to access and utilize: https://github.com/LouisSF/MDT3D

Abstract: Building a multi-modality multi-task neural network toward accurate and robust performance is a de-facto standard in perception task of autonomous driving. However, leveraging such data from multiple sensors to jointly optimize the prediction and planning tasks remains largely unexplored. In this paper, we present FusionAD, to the best of our knowledge, the first unified framework that fuse the information from two most critical sensors, camera and LiDAR, goes beyond perception task. Concretely, we first build a transformer based multi-modality fusion network to effectively produce fusion based features. In constrast to camera-based end-to-end method UniAD, we then establish a fusion aided modality-aware prediction and status-aware planning modules, dubbed FMSPnP that take advantages of multi-modality features. We conduct extensive experiments on commonly used benchmark nuScenes dataset, our FusionAD achieves state-of-the-art performance and surpassing baselines on average 15% on perception tasks like detection and tracking, 10% on occupancy prediction accuracy, reducing prediction error from 0.708 to 0.389 in ADE score and reduces the collision rate from 0.31% to only 0.12%.

Abstract: Computer vision was long a tool used for aiding visually impaired people to move around their environment and avoid obstacles and falls. Solutions are limited to either indoor or outdoor scenes, which limits the kind of places and scenes visually disabled people can be in, including entertainment places such as theatres. Furthermore, most of the proposed computer-vision-based methods rely on RGB benchmarks to train their models resulting in a limited performance due to the absence of the depth modality. In this paper, we propose a novel RGB-D dataset containing theatre scenes with ground truth human actions and dense captions annotations for image captioning and human action recognition: TS-RGBD dataset. It includes three types of data: RGB, depth, and skeleton sequences, captured by Microsoft Kinect. We test image captioning models on our dataset as well as some skeleton-based human action recognition models in order to extend the range of environment types where a visually disabled person can be, by detecting human actions and textually describing appearances of regions of interest in theatre scenes.

Abstract: Multispectral pedestrian detection achieves better visibility in challenging conditions and thus has a broad application in various tasks, for which both the accuracy and computational cost are of paramount importance. Most existing approaches treat RGB and infrared modalities equally, typically adopting two symmetrical CNN backbones for multimodal feature extraction, which ignores the substantial differences between modalities and brings great difficulty for the reduction of the computational cost as well as effective crossmodal fusion. In this work, we propose a novel and efficient framework named WCCNet that is able to differentially extract rich features of different spectra with lower computational complexity and semantically rearranges these features for effective crossmodal fusion. Specifically, the discrete wavelet transform (DWT) allowing fast inference and training speed is embedded to construct a dual-stream backbone for efficient feature extraction. The DWT layers of WCCNet extract frequency components for infrared modality, while the CNN layers extract spatial-domain features for RGB modality. This methodology not only significantly reduces the computational complexity, but also improves the extraction of infrared features to facilitate the subsequent crossmodal fusion. Based on the well extracted features, we elaborately design the crossmodal rearranging fusion module (CMRF), which can mitigate spatial misalignment and merge semantically complementary features of spatially-related local regions to amplify the crossmodal complementary information. We conduct comprehensive evaluations on KAIST and FLIR benchmarks, in which WCCNet outperforms state-of-the-art methods with considerable computational efficiency and competitive accuracy. We also perform the ablation study and analyze thoroughly the impact of different components on the performance of WCCNet.

Abstract: Vision transformers have achieved leading performance on various visual tasks yet still suffer from high computational complexity. The situation deteriorates in dense prediction tasks like semantic segmentation, as high-resolution inputs and outputs usually imply more tokens involved in computations. Directly removing the less attentive tokens has been discussed for the image classification task but can not be extended to semantic segmentation since a dense prediction is required for every patch. To this end, this work introduces a Dynamic Token Pruning (DToP) method based on the early exit of tokens for semantic segmentation. Motivated by the coarse-to-fine segmentation process by humans, we naturally split the widely adopted auxiliary-loss-based network architecture into several stages, where each auxiliary block grades every token's difficulty level. We can finalize the prediction of easy tokens in advance without completing the entire forward pass. Moreover, we keep $k$ highest confidence tokens for each semantic category to uphold the representative context information. Thus, computational complexity will change with the difficulty of the input, akin to the way humans do segmentation. Experiments suggest that the proposed DToP architecture reduces on average $20\% - 35\%$ of computational cost for current semantic segmentation methods based on plain vision transformers without accuracy degradation.

Abstract: Breast cancer is a major cause of cancer death among women, emphasising the importance of early detection for improved treatment outcomes and quality of life. Mammography, the primary diagnostic imaging test, poses challenges due to the high variability and patterns in mammograms. Double reading of mammograms is recommended in many screening programs to improve diagnostic accuracy but increases radiologists' workload. Researchers explore Machine Learning models to support expert decision-making. Stand-alone models have shown comparable or superior performance to radiologists, but some studies note decreased sensitivity with multiple datasets, indicating the need for high generalisation and robustness models. This work devises MammoDG, a novel deep-learning framework for generalisable and reliable analysis of cross-domain multi-center mammography data. MammoDG leverages multi-view mammograms and a novel contrastive mechanism to enhance generalisation capabilities. Extensive validation demonstrates MammoDG's superiority, highlighting the critical importance of domain generalisation for trustworthy mammography analysis in imaging protocol variations.

Abstract: Autonomous navigation in underwater environments presents challenges due to factors such as light absorption and water turbidity, limiting the effectiveness of optical sensors. Sonar systems are commonly used for perception in underwater operations as they are unaffected by these limitations. Traditional computer vision algorithms are less effective when applied to sonar-generated acoustic images, while convolutional neural networks (CNNs) typically require large amounts of labeled training data that are often unavailable or difficult to acquire. To this end, we propose a novel compact deep sonar descriptor pipeline that can generalize to real scenarios while being trained exclusively on synthetic data. Our architecture is based on a ResNet18 back-end and a properly parameterized random Gaussian projection layer, whereas input sonar data is enhanced with standard ad-hoc normalization/prefiltering techniques. A customized synthetic data generation procedure is also presented. The proposed method has been evaluated extensively using both synthetic and publicly available real data, demonstrating its effectiveness compared to state-of-the-art methods.

Abstract: Surveillance videos and images are used for a broad set of applications, ranging from traffic analysis to crime detection. Extrinsic camera calibration data is important for most analysis applications. However, security cameras are susceptible to environmental conditions and small camera movements, resulting in a need for an automated re-calibration method that can account for these varying conditions. In this paper, we present an automated camera-calibration process leveraging a dictionary-based approach that does not require prior knowledge on any camera settings. The method consists of a custom implementation of a Spatial Transformer Network (STN) and a novel topological loss function. Experiments reveal that the proposed method improves the IoU metric by up to 12% w.r.t. a state-of-the-art model across five synthetic datasets and the World Cup 2014 dataset.

Abstract: Accurate 3D tracking of hand and fingers movements poses significant challenges in computer vision. The potential applications span across multiple domains, including human-computer interaction, virtual reality, industry, and medicine. While gesture recognition has achieved remarkable accuracy, quantifying fine movements remains a hurdle, particularly in clinical applications where the assessment of hand dysfunctions and rehabilitation training outcomes necessitate precise measurements. Several novel and lightweight frameworks based on Deep Learning have emerged to address this issue; however, their performance in accurately and reliably measuring fingers movements requires validation against well-established gold standard systems. In this paper, the aim is to validate the handtracking framework implemented by Google MediaPipe Hand (GMH) and an innovative enhanced version, GMH-D, that exploits the depth estimation of an RGB-Depth camera to achieve more accurate tracking of 3D movements. Three dynamic exercises commonly administered by clinicians to assess hand dysfunctions, namely Hand Opening-Closing, Single Finger Tapping and Multiple Finger Tapping are considered. Results demonstrate high temporal and spectral consistency of both frameworks with the gold standard. However, the enhanced GMH-D framework exhibits superior accuracy in spatial measurements compared to the baseline GMH, for both slow and fast movements. Overall, our study contributes to the advancement of hand tracking technology, the establishment of a validation procedure as a good-practice to prove efficacy of deep-learning-based hand-tracking, and proves the effectiveness of GMH-D as a reliable framework for assessing 3D hand movements in clinical applications.

Abstract: Advertising posters, a form of information presentation, combine visual and linguistic modalities. Creating a poster involves multiple steps and necessitates design experience and creativity. This paper introduces AutoPoster, a highly automatic and content-aware system for generating advertising posters. With only product images and titles as inputs, AutoPoster can automatically produce posters of varying sizes through four key stages: image cleaning and retargeting, layout generation, tagline generation, and style attribute prediction. To ensure visual harmony of posters, two content-aware models are incorporated for layout and tagline generation. Moreover, we propose a novel multi-task Style Attribute Predictor (SAP) to jointly predict visual style attributes. Meanwhile, to our knowledge, we propose the first poster generation dataset that includes visual attribute annotations for over 76k posters. Qualitative and quantitative outcomes from user studies and experiments substantiate the efficacy of our system and the aesthetic superiority of the generated posters compared to other poster generation methods.

Abstract: Human motion prediction (HMP) has emerged as a popular research topic due to its diverse applications, but it remains a challenging task due to the stochastic and aperiodic nature of future poses. Traditional methods rely on hand-crafted features and machine learning techniques, which often struggle to model the complex dynamics of human motion. Recent deep learning-based methods have achieved success by learning spatio-temporal representations of motion, but these models often overlook the reliability of motion data. Additionally, the temporal and spatial dependencies of skeleton nodes are distinct. The temporal relationship captures motion information over time, while the spatial relationship describes body structure and the relationships between different nodes. In this paper, we propose a novel spatio-temporal branching network using incremental information for HMP, which decouples the learning of temporal-domain and spatial-domain features, extracts more motion information, and achieves complementary cross-domain knowledge learning through knowledge distillation. Our approach effectively reduces noise interference and provides more expressive information for characterizing motion by separately extracting temporal and spatial features. We evaluate our approach on standard HMP benchmarks and outperform state-of-the-art methods in terms of prediction accuracy.

Abstract: Robust feature matching forms the backbone for most Visual Simultaneous Localization and Mapping (vSLAM), visual odometry, 3D reconstruction, and Structure from Motion (SfM) algorithms. However, recovering feature matches from texture-poor scenes is a major challenge and still remains an open area of research. In this paper, we present a Stereo Visual Odometry (StereoVO) technique based on point and line features which uses a novel feature-matching mechanism based on an Attention Graph Neural Network that is designed to perform well even under adverse weather conditions such as fog, haze, rain, and snow, and dynamic lighting conditions such as nighttime illumination and glare scenarios. We perform experiments on multiple real and synthetic datasets to validate the ability of our method to perform StereoVO under low visibility weather and lighting conditions through robust point and line matches. The results demonstrate that our method achieves more line feature matches than state-of-the-art line matching algorithms, which when complemented with point feature matches perform consistently well in adverse weather and dynamic lighting conditions.

Abstract: Current captioning approaches tend to generate correct but "generic" descriptions that lack real-world knowledge, e.g., named entities and contextual information. Considering that Vision-Language Pre-Training (VLP) models master massive such knowledge from large-scale web-harvested data, it is promising to utilize the generalizability of VLP models to incorporate knowledge into image descriptions. However, using VLP models faces challenges: zero-shot inference suffers from knowledge hallucination that leads to low-quality descriptions, but the generic bias in downstream task fine-tuning hinders the VLP model from expressing knowledge. To address these concerns, we propose a simple yet effective method called Knowledge-guided Replay (K-Replay), which enables the retention of pre-training knowledge during fine-tuning. Our approach consists of two parts: (1) a knowledge prediction task on automatically collected replay exemplars to continuously awaken the VLP model's memory about knowledge, thus preventing the model from collapsing into the generic pattern; (2) a knowledge distillation constraint to improve the faithfulness of generated descriptions hence alleviating the knowledge hallucination. To evaluate knowledge-enhanced descriptions, we construct a novel captioning benchmark KnowCap, containing knowledge of landmarks, famous brands, special foods and movie characters. Experimental results show that our approach effectively incorporates knowledge into descriptions, outperforming strong VLP baseline by 20.9 points (78.7->99.6) in CIDEr score and 20.5 percentage points (34.0%->54.5%) in knowledge recognition accuracy. Our code and data is available at https://github.com/njucckevin/KnowCap.

Abstract: The Class Incremental Semantic Segmentation (CISS) extends the traditional segmentation task by incrementally learning newly added classes. Previous work has introduced generative replay, which involves replaying old class samples generated from a pre-trained GAN, to address the issues of catastrophic forgetting and privacy concerns. However, the generated images lack semantic precision and exhibit out-of-distribution characteristics, resulting in inaccurate masks that further degrade the segmentation performance. To tackle these challenges, we propose DiffusePast, a novel framework featuring a diffusion-based generative replay module that generates semantically accurate images with more reliable masks guided by different instructions (e.g., text prompts or edge maps). Specifically, DiffusePast introduces a dual-generator paradigm, which focuses on generating old class images that align with the distribution of downstream datasets while preserving the structure and layout of the original images, enabling more precise masks. To adapt to the novel visual concepts of newly added classes continuously, we incorporate class-wise token embedding when updating the dual-generator. Moreover, we assign adequate pseudo-labels of old classes to the background pixels in the new step images, further mitigating the forgetting of previously learned knowledge. Through comprehensive experiments, our method demonstrates competitive performance across mainstream benchmarks, striking a better balance between the performance of old and novel classes.

Abstract: This paper introduces a novel approach to leverage the knowledge of existing expert models for training new Convolutional Neural Networks, on domains where task-specific data are limited or unavailable. The presented scheme is applied in offline handwritten signature verification (OffSV) which, akin to other biometric applications, suffers from inherent data limitations due to regulatory restrictions. The proposed Student-Teacher (S-T) configuration utilizes feature-based knowledge distillation (FKD), combining graph-based similarity for local activations with global similarity measures to supervise student's training, using only handwritten text data. Remarkably, the models trained using this technique exhibit comparable, if not superior, performance to the teacher model across three popular signature datasets. More importantly, these results are attained without employing any signatures during the feature extraction training process. This study demonstrates the efficacy of leveraging existing expert models to overcome data scarcity challenges in OffSV and potentially other related domains.

Abstract: Generating visually grounded image captions with specific linguistic styles using unpaired stylistic corpora is a challenging task, especially since we expect stylized captions with a wide variety of stylistic patterns. In this paper, we propose a novel framework to generate Accurate and Diverse Stylized Captions (ADS-Cap). Our ADS-Cap first uses a contrastive learning module to align the image and text features, which unifies paired factual and unpaired stylistic corpora during the training process. A conditional variational auto-encoder is then used to automatically memorize diverse stylistic patterns in latent space and enhance diversity through sampling. We also design a simple but effective recheck module to boost style accuracy by filtering style-specific captions. Experimental results on two widely used stylized image captioning datasets show that regarding consistency with the image, style accuracy and diversity, ADS-Cap achieves outstanding performances compared to various baselines. We finally conduct extensive analyses to understand the effectiveness of our method. Our code is available at https://github.com/njucckevin/ADS-Cap.

Abstract: Change detection (CD) by comparing two bi-temporal images is a crucial task in remote sensing. With the advantages of requiring no cumbersome labeled change information, unsupervised CD has attracted extensive attention in the community. However, existing unsupervised CD approaches rarely consider the seasonal and style differences incurred by the illumination and atmospheric conditions in multi-temporal images. To this end, we propose a change detection with domain shift setting for remote sensing images. Furthermore, we present a novel unsupervised CD method using a light-weight transformer, called UCDFormer. Specifically, a transformer-driven image translation composed of a light-weight transformer and a domain-specific affinity weight is first proposed to mitigate domain shift between two images with real-time efficiency. After image translation, we can generate the difference map between the translated before-event image and the original after-event image. Then, a novel reliable pixel extraction module is proposed to select significantly changed/unchanged pixel positions by fusing the pseudo change maps of fuzzy c-means clustering and adaptive threshold. Finally, a binary change map is obtained based on these selected pixel pairs and a binary classifier. Experimental results on different unsupervised CD tasks with seasonal and style changes demonstrate the effectiveness of the proposed UCDFormer. For example, compared with several other related methods, UCDFormer improves performance on the Kappa coefficient by more than 12\%. In addition, UCDFormer achieves excellent performance for earthquake-induced landslide detection when considering large-scale applications. The code is available at \url{https://github.com/zhu-xlab/UCDFormer}

Abstract: The recently rising markup-to-image generation poses greater challenges as compared to natural image generation, due to its low tolerance for errors as well as the complex sequence and context correlations between markup and rendered image. This paper proposes a novel model named "Contrast-augmented Diffusion Model with Fine-grained Sequence Alignment" (FSA-CDM), which introduces contrastive positive/negative samples into the diffusion model to boost performance for markup-to-image generation. Technically, we design a fine-grained cross-modal alignment module to well explore the sequence similarity between the two modalities for learning robust feature representations. To improve the generalization ability, we propose a contrast-augmented diffusion model to explicitly explore positive and negative samples by maximizing a novel contrastive variational objective, which is mathematically inferred to provide a tighter bound for the model's optimization. Moreover, the context-aware cross attention module is developed to capture the contextual information within markup language during the denoising process, yielding better noise prediction results. Extensive experiments are conducted on four benchmark datasets from different domains, and the experimental results demonstrate the effectiveness of the proposed components in FSA-CDM, significantly exceeding state-of-the-art performance by about 2%-12% DTW improvements. The code will be released at https://github.com/zgj77/FSACDM.

Abstract: We explore a new class of brain encoding model by adding memory-related information as input. Memory is an essential brain mechanism that works alongside visual stimuli. During a vision-memory cognitive task, we found the non-visual brain is largely predictable using previously seen images. Our Memory Encoding Model (Mem) won the Algonauts 2023 visual brain competition even without model ensemble (single model score 66.8, ensemble score 70.8). Our ensemble model without memory input (61.4) can also stand a 3rd place. Furthermore, we observe periodic delayed brain response correlated to 6th-7th prior image, and hippocampus also showed correlated activity timed with this periodicity. We conjuncture that the periodic replay could be related to memory mechanism to enhance the working memory.

Abstract: Driving scene understanding is to obtain comprehensive scene information through the sensor data and provide a basis for downstream tasks, which is indispensable for the safety of self-driving vehicles. Specific perception tasks, such as object detection and scene graph generation, are commonly used. However, the results of these tasks are only equivalent to the characterization of sampling from high-dimensional scene features, which are not sufficient to represent the scenario. In addition, the goal of perception tasks is inconsistent with human driving that just focuses on what may affect the ego-trajectory. Therefore, we propose an end-to-end Interpretable Implicit Driving Scene Understanding (II-DSU) model to extract implicit high-dimensional scene features as scene understanding results guided by a planning module and to validate the plausibility of scene understanding using auxiliary perception tasks for visualization. Experimental results on CARLA benchmarks show that our approach achieves the new state-of-the-art and is able to obtain scene features that embody richer scene information relevant to driving, enabling superior performance of the downstream planning.

Abstract: The learning with noisy labels has been addressed with both discriminative and generative models. Although discriminative models have dominated the field due to their simpler modeling and more efficient computational training processes, generative models offer a more effective means of disentangling clean and noisy labels and improving the estimation of the label transition matrix. However, generative approaches maximize the joint likelihood of noisy labels and data using a complex formulation that only indirectly optimizes the model of interest associating data and clean labels. Additionally, these approaches rely on generative models that are challenging to train and tend to use uninformative clean label priors. In this paper, we propose a new generative noisy-label learning approach that addresses these three issues. First, we propose a new model optimisation that directly associates data and clean labels. Second, the generative model is implicitly estimated using a discriminative model, eliminating the inefficient training of a generative model. Third, we propose a new informative label prior inspired by partial label learning as supervision signal for noisy label learning. Extensive experiments on several noisy-label benchmarks demonstrate that our generative model provides state-of-the-art results while maintaining a similar computational complexity as discriminative models.

Abstract: This paper seeks to address the dense labeling problems where a significant fraction of the dataset can be pruned without sacrificing much accuracy. We observe that, on standard medical image segmentation benchmarks, the loss gradient norm-based metrics of individual training examples applied in image classification fail to identify the important samples. To address this issue, we propose a data pruning method by taking into consideration the training dynamics on target regions using Dynamic Average Dice (DAD) score. To the best of our knowledge, we are among the first to address the data importance in dense labeling tasks in the field of medical image analysis, making the following contributions: (1) investigating the underlying causes with rigorous empirical analysis, and (2) determining effective data pruning approach in dense labeling problems. Our solution can be used as a strong yet simple baseline to select important examples for medical image segmentation with combined data sources.

Abstract: Learning a policy with great generalization to unseen environments remains challenging but critical in visual reinforcement learning. Despite the success of augmentation combination in the supervised learning generalization, naively applying it to visual RL algorithms may damage the training efficiency, suffering from serve performance degradation. In this paper, we first conduct qualitative analysis and illuminate the main causes: (i) high-variance gradient magnitudes and (ii) gradient conflicts existed in various augmentation methods. To alleviate these issues, we propose a general policy gradient optimization framework, named Conflict-aware Gradient Agreement Augmentation (CG2A), and better integrate augmentation combination into visual RL algorithms to address the generalization bias. In particular, CG2A develops a Gradient Agreement Solver to adaptively balance the varying gradient magnitudes, and introduces a Soft Gradient Surgery strategy to alleviate the gradient conflicts. Extensive experiments demonstrate that CG2A significantly improves the generalization performance and sample efficiency of visual RL algorithms.

Abstract: For text-to-video retrieval (T2VR), which aims to retrieve unlabeled videos by ad-hoc textual queries, CLIP-based methods are dominating. Compared to CLIP4Clip which is efficient and compact, the state-of-the-art models tend to compute video-text similarity by fine-grained cross-modal feature interaction and matching, putting their scalability for large-scale T2VR into doubt. For efficient T2VR, we propose TeachCLIP with multi-grained teaching to let a CLIP4Clip based student network learn from more advanced yet computationally heavy models such as X-CLIP, TS2-Net and X-Pool . To improve the student's learning capability, we add an Attentional frame-Feature Aggregation (AFA) block, which by design adds no extra storage/computation overhead at the retrieval stage. While attentive weights produced by AFA are commonly used for combining frame-level features, we propose a novel use of the weights to let them imitate frame-text relevance estimated by the teacher network. As such, AFA provides a fine-grained learning (teaching) channel for the student (teacher). Extensive experiments on multiple public datasets justify the viability of the proposed method.

Abstract: This paper introduces Grounded Image Text Matching with Mismatched Relation (GITM-MR), a novel visual-linguistic joint task that evaluates the relation understanding capabilities of transformer-based pre-trained models. GITM-MR requires a model to first determine if an expression describes an image, then localize referred objects or ground the mismatched parts of the text. We provide a benchmark for evaluating pre-trained models on this task, with a focus on the challenging settings of limited data and out-of-distribution sentence lengths. Our evaluation demonstrates that pre-trained models lack data efficiency and length generalization ability. To address this, we propose the Relation-sensitive Correspondence Reasoning Network (RCRN), which incorporates relation-aware reasoning via bi-directional message propagation guided by language structure. RCRN can be interpreted as a modular program and delivers strong performance in both length generalization and data efficiency.

Abstract: Digital preservation of Cultural Heritage (CH) sites is crucial to protect them against damage from natural disasters or human activities. Creating 3D models of CH sites has become a popular method of digital preservation thanks to advancements in computer vision and photogrammetry. However, the process is time-consuming, expensive, and typically requires specialized equipment and expertise, posing challenges in resource-limited developing countries. Additionally, the lack of an open repository for 3D models hinders research and public engagement with their heritage. To address these issues, we propose Tirtha, a web platform for crowdsourcing images of CH sites and creating their 3D models. Tirtha utilizes state-of-the-art Structure from Motion (SfM) and Multi-View Stereo (MVS) techniques. It is modular, extensible and cost-effective, allowing for the incorporation of new techniques as photogrammetry advances. Tirtha is accessible through a web interface at https://tirtha.niser.ac.in and can be deployed on-premise or in a cloud environment. In our case studies, we demonstrate the pipeline's effectiveness by creating 3D models of temples in Odisha, India, using crowdsourced images. These models are available for viewing, interaction, and download on the Tirtha website. Our work aims to provide a dataset of crowdsourced images and 3D reconstructions for research in computer vision, heritage conservation, and related domains. Overall, Tirtha is a step towards democratizing digital preservation, primarily in resource-limited developing countries.

Abstract: Multi-Object Tracking, also known as Multi-Target Tracking, is a significant area of computer vision that has many uses in a variety of settings. The development of deep learning, which has encouraged researchers to propose more and more work in this direction, has significantly impacted the scientific advancement around the study of tracking as well as many other domains related to computer vision. In fact, all of the solutions that are currently state-of-the-art in the literature and in the tracking industry, are built on top of deep learning methodologies that produce exceptionally good results. Deep learning is enabled thanks to the ever more powerful technology researchers can use to handle the significant computational resources demanded by these models. However, when real-time is a main requirement, developing a tracking system without being constrained by expensive hardware support with enormous computational resources is necessary to widen tracking applications in real-world contexts. To this end, a compromise is to combine powerful deep strategies with more traditional approaches to favor considerably lower processing solutions at the cost of less accurate tracking results even though suitable for real-time domains. Indeed, the present work goes in that direction, proposing a hybrid strategy for real-time multi-target tracking that combines effectively a classical optical flow algorithm with a deep learning architecture, targeted to a human-crowd tracking system exhibiting a desirable trade-off between performance in tracking precision and computational costs. The developed architecture was experimented with different settings, and yielded a MOTA of 0.608 out of the compared state-of-the-art 0.549 results, and about half the running time when introducing the optical flow phase, achieving almost the same performance in terms of accuracy.

Abstract: As a harzard disaster, landslide often brings tremendous losses to humanity, so it's necessary to achieve reliable detection of landslide. However, the problems of visual blur and small-sized dataset cause great challenges for old landslide detection task when using remote sensing data. To reliably extract semantic features, a hyper-pixel-wise contrastive learning augmented segmentation network (HPCL-Net) is proposed, which augments the local salient feature extraction from the boundaries of landslides through HPCL and fuses the heterogeneous infromation in the semantic space from High-Resolution Remote Sensing Images and Digital Elevation Model Data data. For full utilization of the precious samples, a global hyper-pixel-wise sample pair queues-based contrastive learning method, which includes the construction of global queues that store hyper-pixel-wise samples and the updating scheme of a momentum encoder, is developed, reliably enhancing the extraction ability of semantic features. The proposed HPCL-Net is evaluated on a Loess Plateau old landslide dataset and experiment results show that the model greatly improves the reliablity of old landslide detection compared to the previous old landslide segmentation model, where mIoU metric is increased from 0.620 to 0.651, Landslide IoU metric is increased from 0.334 to 0.394 and F1-score metric is increased from 0.501 to 0.565.

Abstract: Long-Term tracking is a hot topic in Computer Vision. In this context, competitive models are presented every year, showing a constant growth rate in performances, mainly measured in standardized protocols as Visual Object Tracking (VOT) and Object Tracking Benchmark (OTB). Fusion-trackers strategy has been applied over last few years for overcoming the known re-detection problem, turning out to be an important breakthrough. Following this approach, this work aims to generalize the fusion concept to an arbitrary number of trackers used as baseline trackers in the pipeline, leveraging a learning phase to better understand how outcomes correlate with each other, even when no target is present. A model and data independence conjecture will be evidenced in the manuscript, yielding a recall of 0.738 on LTB-50 dataset when learning from VOT-LT2022, and 0.619 by reversing the two datasets. In both cases, results are strongly competitive with state-of-the-art and recall turns out to be the first on the podium.

Abstract: As a result of Shadow NeRF and Sat-NeRF, it is possible to take the solar angle into account in a NeRF-based framework for rendering a scene from a novel viewpoint using satellite images for training. Our work extends those contributions and shows how one can make the renderings season-specific. Our main challenge was creating a Neural Radiance Field (NeRF) that could render seasonal features independently of viewing angle and solar angle while still being able to render shadows. We teach our network to render seasonal features by introducing one more input variable -- time of the year. However, the small training datasets typical of satellite imagery can introduce ambiguities in cases where shadows are present in the same location for every image of a particular season. We add additional terms to the loss function to discourage the network from using seasonal features for accounting for shadows. We show the performance of our network on eight Areas of Interest containing images captured by the Maxar WorldView-3 satellite. This evaluation includes tests measuring the ability of our framework to accurately render novel views, generate height maps, predict shadows, and specify seasonal features independently from shadows. Our ablation studies justify the choices made for network design parameters.

Abstract: Motivated by that DETR-based approaches have established new records on COCO detection and segmentation benchmarks, many recent endeavors show increasing interest in how to further improve DETR-based approaches by pre-training the Transformer in a self-supervised manner while keeping the backbone frozen. Some studies already claimed significant improvements in accuracy. In this paper, we take a closer look at their experimental methodology and check if their approaches are still effective on the very recent state-of-the-art such as $\mathcal{H}$-Deformable-DETR. We conduct thorough experiments on COCO object detection tasks to study the influence of the choice of pre-training datasets, localization, and classification target generation schemes. Unfortunately, we find the previous representative self-supervised approach such as DETReg, fails to boost the performance of the strong DETR-based approaches on full data regimes. We further analyze the reasons and find that simply combining a more accurate box predictor and Objects$365$ benchmark can significantly improve the results in follow-up experiments. We demonstrate the effectiveness of our approach by achieving strong object detection results of AP=$59.3\%$ on COCO val set, which surpasses $\mathcal{H}$-Deformable-DETR + Swin-L by +$1.4\%$. Last, we generate a series of synthetic pre-training datasets by combining the very recent image-to-text captioning models (LLaVA) and text-to-image generative models (SDXL). Notably, pre-training on these synthetic datasets leads to notable improvements in object detection performance. Looking ahead, we anticipate substantial advantages through the future expansion of the synthetic pre-training dataset.

Abstract: CLIP, as a foundational vision language model, is widely used in zero-shot image classification due to its ability to understand various visual concepts and natural language descriptions. However, how to fully leverage CLIP's unprecedented human-like understanding capabilities to achieve better zero-shot classification is still an open question. This paper draws inspiration from the human visual perception process: a modern neuroscience view suggests that in classifying an object, humans first infer its class-independent attributes (e.g., background and orientation) which help separate the foreground object from the background, and then make decisions based on this information. Inspired by this, we observe that providing CLIP with contextual attributes improves zero-shot classification and mitigates reliance on spurious features. We also observe that CLIP itself can reasonably infer the attributes from an image. With these observations, we propose a training-free, two-step zero-shot classification method named PerceptionCLIP. Given an image, it first infers contextual attributes (e.g., background) and then performs object classification conditioning on them. Our experiments show that PerceptionCLIP achieves better generalization, group robustness, and better interpretability. For example, PerceptionCLIP with ViT-L/14 improves the worst group accuracy by 16.5% on the Waterbirds dataset and by 3.5% on CelebA.

Abstract: We propose an effective denoising diffusion model for generating high-resolution images (e.g., 1024$\times$512), trained on small-size image patches (e.g., 64$\times$64). We name our algorithm Patch-DM, in which a new feature collage strategy is designed to avoid the boundary artifact when synthesizing large-size images. Feature collage systematically crops and combines partial features of the neighboring patches to predict the features of a shifted image patch, allowing the seamless generation of the entire image due to the overlap in the patch feature space. Patch-DM produces high-quality image synthesis results on our newly collected dataset of nature images (1024$\times$512), as well as on standard benchmarks of smaller sizes (256$\times$256), including LSUN-Bedroom, LSUN-Church, and FFHQ. We compare our method with previous patch-based generation methods and achieve state-of-the-art FID scores on all four datasets. Further, Patch-DM also reduces memory complexity compared to the classic diffusion models.

Abstract: Our approach, which we call Embeddings for Language/Image-aligned X-Rays, or ELIXR, leverages a language-aligned image encoder combined or grafted onto a fixed LLM, PaLM 2, to perform a broad range of tasks. We train this lightweight adapter architecture using images paired with corresponding free-text radiology reports from the MIMIC-CXR dataset. ELIXR achieved state-of-the-art performance on zero-shot chest X-ray (CXR) classification (mean AUC of 0.850 across 13 findings), data-efficient CXR classification (mean AUCs of 0.893 and 0.898 across five findings (atelectasis, cardiomegaly, consolidation, pleural effusion, and pulmonary edema) for 1% (~2,200 images) and 10% (~22,000 images) training data), and semantic search (0.76 normalized discounted cumulative gain (NDCG) across nineteen queries, including perfect retrieval on twelve of them). Compared to existing data-efficient methods including supervised contrastive learning (SupCon), ELIXR required two orders of magnitude less data to reach similar performance. ELIXR also showed promise on CXR vision-language tasks, demonstrating overall accuracies of 58.7% and 62.5% on visual question answering and report quality assurance tasks, respectively. These results suggest that ELIXR is a robust and versatile approach to CXR AI.