WACV 2024 Papers

ENIGMA-51 is a new egocentric dataset acquired in an industrial scenario by 19 subjects who followed instructions to complete the repair of electrical boards using industrial tools (e.g., electric screwdriver) and equipments (e.g., oscilloscope). The 51 egocentric video sequences are densely annotated with a rich set of labels that enable the systematic study of human behavior in the industrial domain. We provide benchmarks on four tasks related to human behavior: 1) untrimmed temporal detection of human-object interactions, 2) egocentric human-object interaction detection, 3) short-term object interaction anticipation and 4) natural language understanding of intents and entities. Baseline results show that the ENIGMA-51 dataset poses a challenging benchmark to study human behavior in industrial scenarios. We publicly release the dataset at https://iplab.dmi.unict.it/ENIGMA-51.

In this paper, we present ECSIC, a novel learned method for stereo image compression. Our proposed method compresses the left and right images in a joint manner by exploiting the mutual information between the images of the stereo image pair using a novel stereo cross attention (SCA) module and two stereo context modules. The SCA module performs cross-attention restricted to the corresponding epipolar lines of the two images and processes them in parallel. The stereo context modules improve the entropy estimation of the second encoded image by using the first image as a context. We conduct an extensive ablation study demonstrating the effectiveness of the proposed modules and a comprehensive quantitative and qualitative comparison with existing methods. ECSIC achieves state-of-the-art performance in stereo image compression on the two popular stereo image datasets Cityscapes and InStereo2k while allowing for fast encoding and decoding.

Multimodal sentiment analysis (MSA) and multimodal emotion recognition (MER) tasks have gained a surge of attention in recent years. Although both tasks share common ground in many ways, they are often treated as a separate task. In this work, we propose, EASUM, a new training scheme for bridging the MSA and MER tasks. EASUM aims to bring mutual benefits to both tasks based on the premise that the sentiment and emotion are closely related; hence each information should provide deeper insight into one's affective state to complement the other. We exploit this premise to further improve the performance of each task by 1) first training a domain general model using four benchmark datasets from the MSA and MER tasks: CMU-MOSI, CMU-MOSEI, MELD, and IEMOCAP. Depending on the dataset, the domain general model learns to predict sentiment or emotion values based on the domain invariant features. 2) Then these values are later used as auxiliary pseudo labels when training a domain specific model for each task. Our premise as well as new training scheme are validated through extensive experiments on the four benchmark datasets. The results also demonstrate that the proposed method outperforms the state-of-the-art on the CMU-MOSI, CMU-MOSEI, and MELD datasets, and performs comparable to the state-of-the-art on the IEMOCAP dataset while using approximately 40% fewer parameters.

Vision transformers (ViT) usually extract features via forwarding all the tokens in the self-attention layers from top to toe. In this paper, we introduce dynamic token-pass vision transformers (DoViT) for semantic segmentation, which can adaptively reduce the inference cost for images with different complexity. DoViT gradually stops partial easy tokens from self-attention calculation and keeps the hard tokens forwarding until meeting the stopping criteria. We employ lightweight auxiliary heads to make the token-pass decision and divide the tokens into keeping/stopping parts. With a token separate calculation, the self-attention layers are speeded up with sparse tokens and still work friendly with hardware. A token reconstruction module is built to collect and reset the grouped tokens to their original position in the sequence, which is necessary to predict correct semantic masks. We conduct extensive experiments on two common semantic segmentation tasks, and demonstrate that our method greatly reduces about 40% 60% FLOPs and the drop of mIoU is within 0.8% for various segmentation transformers. The throughput and inference speed of ViT-L/B are increased to more than 2x on Cityscapes.

Effectively leveraging multimodal data such as various images, laboratory tests and clinical information is gaining traction in a variety of AI-based medical diagnosis and prognosis tasks. Most existing multi-modal techniques only focus on enhancing their performance by leveraging the differences or shared features from various modalities and fusing feature across different modalities. These approaches are generally not optimal for clinical settings, which pose the additional challenges of limited training data, as well as being rife with redundant data or noisy modality channels, leading to subpar performance. To address this gap, we study the robustness of existing methods to data redundancy and noise and propose a generalized dynamic multimodal information bottleneck framework for attaining a robust fused feature representation. Specifically, our information bottleneck module serves to filter out the task-irrelevant information and noises in the fused feature, and we further introduce a sufficiency loss to prevent dropping of task-relevant information, thus explicitly preserving the sufficiency of prediction information in the distilled feature. We validate our model on an in-house and a public COVID-19 dataset for mortality prediction as well as two public biomedical datasets for diagnostic tasks. Extensive experiments show that our method surpasses the state-of-the-art and is significantly more robust, being the only method to remain performant when large-scale noisy channels exist. Our code is publicly available at https://github.com/Anonymous-PaperSubmission/DMIB.

Previous methods to solve the problem of metal artifact reduction (MAR) have mostly focused on 2D MAR, making it challenging to apply to problems with 3-dimensional CT such as CBCT. In this paper, we propose a novel approach for 3D MAR which utilizes two diffusion models to model the metal-free CBCT prior and metal artifact prior. Through dual-domain guidance in the image and projection domains, the 3D connectivity is enhanced in the restored images. Moreover, we propose a memory-efficient technique for an efficient sampling of 3-dimensional data, which reduces the memory usage by orders of magnitude. Experiments show that our method achieves the state-of-the-art performance not only with synthetic data but also with real-world clinical and out-of-distribution data.

Concept bottleneck models have been successfully used for explainable machine learning by encoding information within the model with a set of human-defined concepts. In the context of human-assisted or autonomous driving, explainability models can help user acceptance and understanding of decisions made by the autonomous vehicle, which can be used to rationalize and explain driver or vehicle behavior. We propose a new approach using concept bottlenecks as visual features for control command predictions and explanations of user and vehicle behavior. We learn a human understandable concept layer that we use to explain sequential driving scenes while learning vehicle control commands. This approach can then be used to determine whether a change in a preferred gap or steering commands from a human (or autonomous vehicle) is led by an external stimulus or change in preferences. We achieve competitive performance to latent visual features while gaining interpretability within our model setup.

Data augmentation is one of the most powerful techniques for regularizing deep learning models and improving their recognition performance in a variety of tasks and domains. However, this holds for standard in-domain settings, in which the training and test data follow the same distribution. For the out-domain, in which the test data follows a different and unknown distribution, the best recipe for data augmentation is not clear. In this paper, we show that also for out-domain or domain generalization settings, data augmentation can bring a conspicuous and robust improvement in performance. For doing that, we propose a simple procedure: i) use uniform sampling on standard data augmentation transformations ii) increase transformations strength to adapt to the higher data variance expected when working out of domain iii) devise a new reward function to reject extreme transformations that can harm the training. With this simple formula, our data augmentation scheme achieves comparable or better results to state-of-the-art performance on most domain generalization datasets.

Domain generalization (DG) approaches intend to extract domain invariant features that can lead to a more robust deep learning model. In this regard, style augmentation is a strong DG method taking advantage of instance-specific feature statistics containing informative style characteristics to synthetic novel domains. While it is one of the state-of-the-art methods, prior works on style augmentation have either disregarded the interdependence amongst distinct feature channels or have solely constrained style augmentation to linear interpolation. To address these research gaps, in this work, we introduce a novel augmentation approach, named Correlated Style Uncertainty (CSU), surpassing the limitations of linear interpolation in style statistic space and simultaneously preserving vital correlation information. Our method's efficacy is established through extensive experimentation on diverse cross-domain computer vision and medical imaging classification tasks: PACS, Office-Home, and Camelyon17 datasets, and the Duke-Market1501 instance retrieval task. The results showcase a remarkable improvement margin over existing state-of-the-art techniques. The source code is available: https://github.com/freshman97/CSU.

We propose a novel approach for domain generalisation (DG) leveraging risk distributions to characterise domains, thereby achieving domain invariance. In our findings, risk distributions effectively highlight differences between training domains and reveal their inherent complexities. In testing, we may observe similar, or potentially intensifying in magnitude, divergences between risk distributions. Hence, we propose a compelling proposition: Minimising the divergences between risk distributions across training domains leads to robust invariance for DG. The key rationale behind this concept is that a model, trained on domain-invariant or stable features, may consistently produce similar risk distributions across various domains. Building upon this idea, we propose Risk Distribution Matching (RDM). Using the maximum mean discrepancy (MMD) distance, RDM aims to minimise the variance of risk distributions across training domains. However, when the number of domains increases, the direct optimisation of variance leads to linear growth in MMD computations, resulting in inefficiency. Instead, we propose an approximation that requires only one MMD computation, by aligning just two distributions: that of the worst-case domain and the aggregated distribution from all domains. Notably, this method empirically outperforms optimising distributional variance while being computationally more efficient. Unlike conventional DG matching algorithms, RDM stands out for its enhanced efficacy by concentrating on scalar risk distributions, sidestepping the pitfalls of high-dimensional challenges seen in feature or gradient matching. Our extensive experiments on standard benchmark datasets demonstrate that RDM shows superior generalisation capability over state-of-the-art DG methods.

Large vision-language representation learning models like CLIP have demonstrated impressive performance for zero-shot transfer to downstream tasks while largely benefiting from inter-modal (image-text) alignment via contrastive objectives. This downstream performance can further be enhanced by full-scale fine-tuning which is often compute intensive, requires large labelled data, and can reduce out-of-distribution (OOD) robustness. Furthermore, sole reliance on inter-modal alignment might overlook the rich information embedded within each individual modality. In this work, we introduce a sample-efficient domain adaptation strategy for CLIP, termed Domain Aligned CLIP (DAC), which improves both intra-modal (image-image) and inter-modal alignment on target distributions without fine-tuning the main model. For intra-modal alignment, we introduce a lightweight adapter that is specifically trained with an intra-modal contrastive objective. To improve inter-modal alignment, we introduce a simple framework to modulate the precomputed class text embeddings. The proposed few-shot fine-tuning framework is computationally efficient, robust to distribution shifts, and does not alter CLIP's parameters. We study the effectiveness of DAC by benchmarking on 11 widely used image classification tasks with consistent improvements in 16-shot classification upon strong baselines by about 2.3% and demonstrate competitive performance on 4 OOD robustness benchmarks.

In this work, we address the novel and challenging problem of domain adaptive 3D shape retrieval from single 2D images (DA-IBSR). While the existing image-based 3D shape retrieval (IBSR) problem focuses on modality alignment for retrieving a matchable 3D shape from a shape repository given a 2D image query, it does not consider any distribution shift between the training and testing image-shape pairs, making the performance of off-the-shelves IBSR methods subpar. In contrast, the proposed DA-IBSR addresses the non-trivial problem of modality shift as well distribution shift across training and test sets. To address these issues, we propose an end-to-end trainable model called DAIS-NET. Our objective is to align the images and shapes separately from both domains while simultaneously learn a shared embedding space for the 2D and 3D modalities. The former problem is addressed by separately employing maximum mean discrepancy loss across the 2D images and 3D shapes of the two domains. To address the modality alignment, we incorporate the notion of negative sample mining and employ triplet loss to bridge the gap between positive 2D-3D pairs (of same class) and increase the separation between negative 2D-3D pairs (of different class). Additionally, we employ an entropy minimization strategy to align the unlabeled target domain data in the semantic space. To evaluate our proposed approach, we define the experimental setting of DA-IBSR on the following benchmarks: SHREC'14 <-> Pix3D and ShapeNet <-> SHREC'14. Considering the novelty of the problem statement, we have demonstrated that the issue of domain gap is prevalent by comparing our method with the existing literature. Additionally, through extensive evaluations, we demonstrate the capability of DAIS-NET to successfully mitigate this domain gap in image based 3D shape retrieval.

Document image dewarping is a crucial task in computer vision with numerous practical applications. The control point method, as a popular image dewarping approach, has attracted attention due to its simplicity and efficiency. However, inaccurate control point prediction due to varying background noises and deformation types can result in unsatisfactory performance. To address these issues, we propose a robust document dewarping approach for real-life images, namely DocReal, which utilizes Enet to effectively remove background noise and an attention-enhanced control point (AECP) module to better capture local deformations. Moreover, we augment the training data by synthesizing 2D images with 3D deformations and additional deformation types. Our proposed method achieves state-of-the-art performance on the DocUNet benchmark and a newly proposed benchmark of 200 Chinese distorted images, exhibiting superior dewarping accuracy, OCR performance, and robustness to various types of image distortion.

Object detectors are at the heart of many semi- and fully autonomous decision systems and are poised to become even more indispensable. They are, however, still lacking in accessibility and can sometimes produce unreliable predictions. Especially concerning in this regard are the (essentially hand-crafted) non-maximum suppression algorithms that lead to an obfuscated prediction process and biased confidence estimates. We show that we can eliminate classic NMS-style post-processing by using IoU-aware calibration. IoU-aware calibration is a conditional Beta calibration; this makes it parallelizable with no hyper-parameters. Instead of arbitrary cutoffs or discounts, it implicitly accounts for the likelihood of each detection being a duplicate and adjusts the confidence score accordingly, resulting in empirically based precision estimates for each detection. Our extensive experiments on diverse detection architectures show that the proposed IoU-aware calibration can successfully model duplicate detections and improve calibration. Compared to the standard sequential NMS and calibration approach, our joint modeling can deliver performance gains over the best NMS-based alternative while producing consistently better-calibrated confidence predictions with less complexity.

The Lip Reading Sentences-3 (LRS3) benchmark has primarily been the focus of intense research in visual speech recognition (VSR) during the last few years. As a result, there is an increased risk of overfitting to its excessively used test set, which is only one hour duration. To alleviate this issue, we build a new VSR test set by closely following the LRS3 dataset creation processes. We then evaluate and analyse the extent to which the current VSR models generalize to the new test data. We evaluate a broad range of publicly available VSR models and find significant drops in performance on our test set, compared to their corresponding LRS3 results. Our results suggest that the increase in word error rates is caused by the models' inability to generalize to slightly "harder" and more realistic lip sequences than those found in the LRS3 test set. Our new test benchmark will be made public in order to enable future research towards more robust VSR models.

A commonly accepted hypothesis is that models with higher accuracy on Imagenet perform better on other downstream tasks, leading to much research dedicated to optimizing Imagenet accuracy. Recently this hypothesis has been challenged by evidence showing that self-supervised models transfer better than their supervised counterparts, despite their inferior Imagenet accuracy. This calls for identifying the additional factors, on top of Imagenet accuracy, that make models transferable. In this work we show that high diversity of the filters learnt by the model promotes transferability jointly with Imagenet accuracy. Encouraged by the recent transferability results of self-supervised models, we use a simple procedure to combine self-supervised and supervised pretraining and generate models with both high diversity and high accuracy, and as a result high transferability. We experiment with several architectures and multiple downstream tasks, including both single-label and multi-label classification.

Self-supervised learning is well known for its remarkable performance in representation learning and various downstream computer vision tasks. Recently, Positive-pair-Only Contrastive Learning (POCL) has achieved reliable performance without the need to construct positive-negative training sets. It reduces memory requirements by lessening the dependency on the batch size. The POCL method typically uses a single objective function to extract the distortion invariant representation (DIR), which describes the proximity of positive-pair representations affected by different distortions. This objective function implicitly enables the model to filter out or ignore the distortion variant representation (DVR) affected by different distortions. However, some recent studies have shown that proper use of DVR in contrastive can optimize the performance of models in some downstream domain-specific tasks. In addition, these POCL methods have been observed to be sensitive to augmentation strategies. To address these limitations, we propose a novel POCL framework named Distortion-Disentangled Contrastive Learning (DDCL) and a Distortion-Disentangled Loss (DDL). Our approach is the first to explicitly and adaptively disentangle and exploit the DVR inside the model and feature stream to improve the overall representation utilization efficiency, robustness, and representation ability. Experiments demonstrate our framework's superiority to Barlow Twins and Simsiam in terms of convergence, representation quality (Including transferability and generality), and robustness on several benchmark datasets.

Image matting requires high-quality pixel-level human annotations to support the training of a deep model in recent literature. Whereas such annotation is costly and hard to scale, significantly holding back the development of the research. In this work, we make the first attempt towards addressing this problem, by proposing a self-supervised pre-training approach that can leverage infinite numbers of data to boost the matting performance. The pre-training task is designed in a similar manner as image matting, where random trimap and alpha matte are generated to achieve an image disentanglement objective. The pre-trained model is then used as an initialisation of the downstream matting task for fine-tuning. Extensive experimental evaluations show that the proposed approach outperforms both the state-of-the-art matting methods and other alternative self-supervised initialisation approaches by a large margin. We also show the robustness of the proposed approach over different backbone architectures. Our project page is available at https://crystraldo.github.io/dpt_mat/.

In this paper, a discriminator-free adversarial-based Unsupervised Domain Adaptation (UDA) for Multi-Label Image Classification (MLIC) referred to as DDA-MLIC is proposed. Recently, some attempts have been made for introducing adversarial-based UDA methods in the context of MLIC. However, these methods which rely on an additional discriminator subnet present one major shortcoming. The learning of domain-invariant features may harm their task-specific discriminative power, since the classification and discrimination tasks are decoupled. Herein, we propose to overcome this issue by introducing a novel adversarial critic that is directly deduced from the task-specific classifier. Specifically, a two-component Gaussian Mixture Model (GMM) is fitted on the source and target predictions in order to distinguish between two clusters. This allows extracting a Gaussian distribution for each component. The resulting Gaussian distributions are then used for formulating an adversarial loss based on a Frechet distance. The proposed method is evaluated on several multi-label image datasets covering three different types of domain shift. The obtained results demonstrate that DDA-MLIC outperforms existing state-of-the-art methods in terms of precision while requiring a lower number of parameters. The code is publicly available at github.com/cvi2snt/DDA-MLIC.