WACV 2024 Papers

Temporal action segmentation relates to high-level video understanding, commonly formulated as frame-wise classification of untrimmed videos into predefined actions. Fully-supervised deep-learning approaches require dense video annotations which are time and money consuming. Furthermore, the temporal boundaries between consecutive actions typically are not well-defined, leading to inherent ambiguity and inter-rater disagreement. A promising approach to remedy these limitations is timestamp supervision, requiring only one labeled frame per action instance in a training video. In this work, we reformulate the task of temporal segmentation as a graph segmentation problem with weakly-labeled vertices. We introduce an efficient segmentation method based on random walks on graphs, obtained by solving a sparse system of linear equations. Furthermore, the proposed technique can be employed in any one or combination of the following forms: (1) as a standalone solution for generating dense pseudo-labels from timestamps; (2) as a training loss; (3) as a smoothing mechanism given intermediate predictions. Extensive experiments with three datasets (50Salads, Breakfast, GTEA) show that our method competes with state-of-the-art, and allows the identification of regions of uncertainty around action boundaries.

In semantic navigation, a top-down map with accurate and complete semantic information is vital to subsequent decision-making. However, due to occlusions and limitations of the robot's field of view (FOV), there are often unobserved areas in the top-down maps. To address this problem, recent works have studied semantic map prediction to complete the top-down maps. In this work, we propose to improve map prediction by integrating relational information. We propose RSMPNet, a relationship-guided semantic map prediction network, which makes use of semantic and spatial relationships to predict unobserved areas from accumulated semantic maps. Specifically, we propose a Relationship Reasoning Layer that includes two modules, namely 1) the Semantic Relationship Graph Reasoning Module (SeGRM) to capture the semantic relationship and 2) the Spatial Relationship Graph Reasoning Module (SpGRM) to utilize the spatial relationship. We also design a semantic relationship enhanced loss to enhance our model to learn semantic relationship information. Experiments show the effectiveness of our proposed network which achieves state-of-the-art performance in semantic map prediction. Our code and datasets are publicly available at https://github.com/jws39/semantic-mapprediction

Effectively capturing intricate interactions among road users is of critical importance to achieving safe navigation for autonomous vehicles. While graph learning (GL) has emerged as a promising approach to tackle this challenge, existing GL models rely on predefined domain-specific graph extraction rules that often fail in real-world drastically changing scenarios. Additionally, these graph extraction rules severely impede the capability of existing GL methods to generalize knowledge across domains. To address this issue, we propose RoadScene2Graph (RS2G), an innovative autonomous scenario understanding framework with a novel data-driven graph extraction and modeling approach that dynamically captures the diverse relations among road users. Our evaluations demonstrate that on average RS2G outperforms the state-of-the-art (SOTA) rule-based graph extraction method by 4.47% and the SOTA deep learning model by 22.19% in subjective risk assessment. More importantly, RS2G delivers notably better performance in transferring knowledge gained from simulation environments to unseen real-world scenarios.

Deep learning (DL) networks have achieved remarkable performance in infrared small target detection (ISTD). However, these structures exhibit a deficiency in interpretability and are widely regarded as black boxes, as they disregard domain knowledge in ISTD. To alleviate this issue, this work proposes an interpretable deep network for detecting infrared dim targets, dubbed RPCANet. Specifically, our approach formulates the ISTD task as sparse target extraction, low-rank background estimation, and image reconstruction in a relaxed Robust Principle Component Analysis (RPCA) model. By unfolding the iterative optimization updating steps into a deep-learning framework, time-consuming and complex matrix calculations are replaced by theory-guided neural networks. RPCANet detects targets with clear interpretability and preserves the intrinsic image feature, instead of directly transforming the detection task into a matrix decomposition problem. Extensive experiments substantiate the effectiveness of our deep unfolding framework and demonstrate its trustworthy results, surpassing baseline methods in both qualitative and quantitative evaluations.

Facial expression recognition (FER) has greatly benefited from deep learning but still faces challenges in dataset collection due to the nuanced nature of facial expressions. In this study, we present a novel unlabeled dataset and semi-supervised contrastive learning framework that utilizes Reaction Mashup (RM) videos, a video that includes multiple individuals reacting to the same film. We created a Reaction Mashup dataset (RMset) from these videos. Our framework integrates three distinct modules: A classification module for supervised facial expression categorization, an attention module for inter-sample attention learning, and a contrastive module for attention-based contrastive learning using RMset. We utilize both the classification and attention modules for the initial training, subsequently incorporating the contrastive module to enhance the learning process. Our experiments demonstrate that our method improves feature learning and outperforms state-of-the-art models on three benchmark FER datasets. Codes are available at https://github.com/yunseongcho/RMFER.

Shape analysis tasks, including mesh classification, segmentation, and retrieval demonstrate symmetries in Euclidean space and should be invariant to geometric transformations such as rotation and translation. However, existing methods in mesh analysis often rely on extensive data augmentation and more complex analysis models to handle 3D rotations. Despite these efforts, rotation invariance is not guaranteed, which can significantly reduce accuracy when test samples undergo arbitrary rotations, because the analysis method struggles to generalize to the unknown orientations of the test samples. To address these challenges, our work presents a novel approach that employs graph neural networks (GNNs) to analyze mesh-structured data. Our proposed GNN layer, aggregation function, and local pooling layer are equivariant to the rotation, reflection and translation of 3D shapes, making them suitable building blocks for our proposed rotation-invariant network for the classification of mesh models. Therefore, our proposed approach does not need rotation augmentation, and we can maintain accuracy even when test samples undergo arbitrary rotations. Extensive experiments on various datasets demonstrate that our methods achieve state-of-the-art performance.

While there exists a great deal of labeled in-the-wild human data, the same is not true for animals. Manually creating new labels for the full range of animal species would take years of effort from the community. We are also now seeing the emerging potential for computer vision methods in areas like animal conservation, which is an additional motivation for this direction of research. Key to our approach is the ability to easily generate as many labeled training images as we desire across a range of different modalities. To achieve this, we present a new large scale canine motion capture dataset and parametric canine body and texture model. These are used to produce the first large scale, multi-domain, multi-task dataset for canine body analysis comprising of detailed synthetic labels on both real images and fully synthetic images in a range of realistic poses. We also introduce the first pose prior for animals in the form of a variational pose prior for canines which is used to fit the parametric model to images of canines. We demonstrate the effectiveness of our labels for training computer vision models on tasks such as parts-based segmentation and pose estimation and show such models can generalise to other animal species without additional training.

Multimodal deep sensor fusion has the potential to enable autonomous vehicles to visually understand their surrounding environments in all weather conditions. However, existing deep sensor fusion methods usually employ convoluted architectures with intermingled multimodal features, requiring large coregistered multimodal datasets for training. In this work, we present an efficient and modular RGB-X fusion network that can leverage and fuse pretrained single-modal models via scene-specific fusion modules, thereby enabling joint input-adaptive network architectures to be created using small, coregistered multimodal datasets. Our experiments demonstrate the superiority of our method compared to existing works on RGB-thermal and RGB-gated datasets, performing fusion using only a small amount of additional parameters. Our code is available at https://github.com/dsriaditya999/RGBXFusion.

The widespread adoption of Neural Radiance Fields (NeRFs) have ensured significant advances in the domain of novel view synthesis in recent years. These models capture a volumetric radiance field of a scene, creating highly convincing, dense, photorealistic models through the use of simple, differentiable rendering equations. Despite their popularity, these algorithms suffer from severe ambiguities in visual data inherent to the RGB sensor, which means that although images generated with view synthesis can visually appear very believable, the underlying 3D model will often be wrong. This considerably limits the usefulness of these models in practical applications like Robotics and Extended Reality (XR), where an accurate dense 3D reconstruction otherwise would be of significant value. In this paper, we present the vital differences between view synthesis models and 3D reconstruction models. We also comment on why a depth sensor is essential for modeling accurate geometry in general outward-facing scenes using the current paradigm of novel view synthesis methods. Focusing on the structure-from-motion task, we practically demonstrate this need by extending the Plenoxel radiance field model: Presenting an analytical differential approach for dense mapping and tracking with radiance fields based on RGB-D data without a neural network. Our method achieves state-of-the-art results in both mapping and tracking tasks, while also being faster than competing neural network-based approaches. The code is available at: https://github.com/ysus33/RGB-D_Plenoxel_Mapping_Tracking.git

Fully-test-time adaptation (F-TTA) can mitigate performance loss due to distribution shifts between train and test data (1) without access to the training data, and (2) without knowledge of the model training procedure. In online F-TTA, a pre-trained model is adapted using a stream of test samples by minimizing a self-supervised objective, such as entropy minimization. However, models adapted with online using entropy minimization, are unstable especially in single sample settings, leading to degenerate solutions, and limiting the adoption of TTA inference strategies. Prior works identify noisy, or unreliable, samples as a cause of failure in online F-TTA. One solution is to ignore these samples, which can lead to bias in the update procedure, slow adaptation, and poor generalization. In this work, we present a general framework for improving robustness of F-TTA to these noisy samples, inspired by self-paced learning and robust loss functions. Our proposed approach, Robust Entropy Adaptive Loss Minimization (REALM), achieves better adaptation accuracy than previous approaches throughout the adaptation process on corruptions of CIFAR-10 and ImageNet-1K, demonstrating its effectiveness.

One of the most challenging problems in audio-driven talking head generation is achieving high-fidelity detail while ensuring precise synchronization. Given only a single reference image, extracting meaningful identity attributes becomes even more challenging, often causing the network to mirror the facial and lip structures too closely. To address these issues, we introduce RADIO, a framework engineered to yield high-quality dubbed videos regardless of the pose or expression in reference images. The key is to modulate the decoder layers using latent space composed of audio and reference features. Additionally, we incorporate ViT blocks into the decoder to emphasize high-fidelity details, especially in the lip region. Our experimental results demonstrate that RADIO displays high synchronization without the loss of fidelity. Especially in harsh scenarios where the reference frame deviates significantly from the ground truth, our method outperforms state-of-the-art methods, highlighting its robustness.

In this study, we explore sketch-based object localization on natural images. Given a crude hand-drawn object sketch, the task is to locate all instances of that object in the target image. This problem proves difficult due to the abstract nature of hand-drawn sketches, variations in the style and quality of sketches, and the large domain gap between the sketches and the natural images. Existing solutions address this using attention-based frameworks to merge query information into image features. Yet, these methods often integrate query features after independently learning image features, causing inadequate alignment and as a result incorrect localization. In contrast, we propose a novel sketch-guided vision transformer encoder that uses cross-attention after each block of the transformer-based image encoder to learn query-conditioned image features, leading to stronger alignment with the query sketch. Further, at the decoder's output, object and sketch features are refined better to align the representation of objects with the sketch query, thereby improving localization. The proposed model also generalizes to the object categories not seen during training, as the target image features learned by the proposed model are query-aware. Our framework can utilize multiple sketch queries via a trainable novel sketch fusion strategy. The model is evaluated on the images from the public benchmark, MS-COCO, using the sketch queries from QuickDraw! and Sketchy datasets. Compared with existing localization methods, the proposed approach gives a 6.6% and 8.0% improvement in mAP for seen objects using sketch queries from QuickDraw! and Sketchy datasets, respectively, and a 12.2% improvement in AP@50 for large objects that are 'unseen' during training.

Psychological trait estimation from external factors such as movement and appearance is a challenging and long-standing problem in psychology, and is principally based on the psychological theory of embodiment. To date, attempts to tackle this problem have utilized private small-scale datasets with intrusive body-attached sensors. Potential applications of an automated system for psychological trait estimation include estimation of occupational fatigue and psychology, and marketing and advertisement. In this work, we propose PsyMo (Psychological traits from Motion), a novel, multi-purpose and multi-modal dataset for exploring psychological cues manifested in walking patterns. We gathered walking sequences from 312 subjects in 7 different walking variations and 6 camera angles. In conjunction with walking sequences, participants filled in 6 psychological questionnaires, totaling 17 psychometric attributes related to personality, self-esteem, fatigue, aggressiveness and mental health. We propose two evaluation protocols for psychological trait estimation. Alongside the estimation of self-reported psychological traits from gait, the dataset can be used as a drop-in replacement to benchmark methods for gait recognition. We anonymize all cues related to the identity of the subjects and publicly release only silhouettes, 2D / 3D human skeletons and 3D SMPL human meshes.

Most pruning methods focus on determining redundant channels from the pre-trained model. However, they overlook the cost of training large networks and the significance of selecting channels for effective reconfiguration. In this paper, we present a "pruning from scratch" framework that considers reconfiguration and expression capacity. Our Shared Pruning Module (SPM) handles a channel alignment problem in residual blocks for lossless reconfiguration after pruning. Moreover, we introduce nuclear norm-based regularization to preserve the representability of large networks during the pruning process. By combining it with MACs-based regularization, we achieve an efficient and powerful pruned network while compressing towards target MACs. The experimental results demonstrate that our method prunes redundant channels effectively to enhance representation capacity of the network. Our approach compresses ResNet50 on ImageNet without requiring additional resources, achieving a top-1 accuracy of 75.25% with only 41% of the original model's MACs. Code is available at https://github.com/jsleeg98/NuSPM.

DDIM inversion has revealed the remarkable potential of real image editing within diffusion-based methods. However, the accuracy of DDIM reconstruction degrades as larger classifier-free guidance (CFG) scales being used for enhanced editing. Null-text inversion (NTI) optimizes null embeddings to align the reconstruction and inversion trajectories with larger CFG scales, enabling real image editing with cross-attention control. Negative-prompt inversion (NPI) further offers a training-free closed-form solution of NTI. However, it may introduce artifacts and is still constrained by DDIM reconstruction quality. To overcome these limitations, we propose proximal guidance and incorporate it to NPI with cross-attention control. We enhance NPI with a regularization term and inversion guidance, which reduces artifacts while capitalizing on its training-free nature. Additionally, we extend the concepts to incorporate mutual self-attention control, enabling geometry and layout alterations in the editing process. Our method provides an efficient and straightforward approach, effectively addressing real image editing tasks with minimal computational overhead.

Binary segmentation is the main task underpinning several remote sensing applications, which are particularly interested in identifying and monitoring a specific category/object. Although extremely important, such a task has several challenges, including huge intra-class variance for the background and data imbalance. Furthermore, most works tackling this task partially or completely ignore one or both of these challenges and their developments. In this paper, we propose a novel method to perform imbalanced binary segmentation of remote sensing images based on deep networks, prototypes, and contrastive loss. The proposed approach allows the model to focus on learning the foreground class while alleviating the class imbalance problem by allowing it to concentrate on the most difficult background examples. The results demonstrate that the proposed method outperforms state-of-the-art techniques for imbalanced binary segmentation of remote sensing images while taking much less training time.

The lack of explainability of deep learning models limits the adoption of such models in clinical practice. Prototype-based models can provide inherent explainable predictions, but these have predominantly been designed for classification tasks, despite many important tasks in medical imaging being continuous regression problems. Therefore, in this work, we present ExPeRT: an explainable prototype-based model specifically designed for regression tasks. Our proposed model makes a sample prediction from the distances to a set of learned prototypes in latent space, using a weighted mean of prototype labels. The distances in latent space are regularized to be relative to label differences, and each of the prototypes can be visualized as a sample from the training set. The image-level distances are further constructed from patch-level distances, in which the patches of both images are structurally matched using optimal transport. This thus provides an example-based explanation with patch-level detail at inference time. We demonstrate our proposed model for brain age prediction on two imaging datasets: adult MR and fetal ultrasound. Our approach achieved state-of-the-art prediction performance while providing insight into the model's reasoning process.

Action recognition models have achieved impressive results by incorporating scene-level annotations, such as objects, their relations, 3D structure, and more. However, obtaining annotations of scene structure for videos requires a significant amount of effort to gather and annotate, making these methods expensive to train. In contrast, synthetic datasets generated by graphics engines provide powerful alternatives for generating scene-level annotations across multiple tasks. In this work, we propose an approach to leverage synthetic scene data for improving video understanding. We present a multi-task prompt learning approach for video transformers, where a shared video transformer backbone is enhanced by a small set of specialized parameters for each task. Specifically, we add a set of "task prompts", each corresponding to a different task, and let each prompt predict task-related annotations. This design allows the model to capture information shared among synthetic scene tasks as well as information shared between synthetic scene tasks and a real video downstream task throughout the entire network. We refer to this approach as "Promptonomy", since the prompts model task-related structure. We propose the PromptonomyViT model (PViT), a video transformer that incorporates various types of scene-level information from synthetic data using the "Promptonomy" approach. PViT shows strong performance improvements on multiple video understanding tasks and datasets. Project page: https://ofir1080.github.io/PromptonomyViT/

Recently, CLIP-based approaches have exhibited remarkable performance on generalization and few-shot learning tasks, fueled by the power of contrastive language-vision pre-training. In particular, prompt tuning has emerged as an effective strategy to adapt the pre-trained language-vision models to downstream tasks by employing task-related textual tokens. Motivated by this progress, in this work we question whether other fundamental problems, such as weakly supervised semantic segmentation (WSSS), can benefit from prompt tuning. Our findings reveal two interesting observations that shed light on the impact of prompt tuning on WSSS. First, modifying only the class token of the text prompt results in a greater impact on the Class Activation Map (CAM), compared to arguably more complex strategies that optimize the context. And second, the class token associated with the image ground truth does not necessarily correspond to the category that yields the best CAM. Motivated by these observations, we introduce a novel approach based on a PrOmpt cLass lEarning (POLE) strategy. Through extensive experiments we demonstrate that our simple, yet efficient approach achieves SOTA performance in a well-known WSSS benchmark. These results highlight not only the benefits of language-vision models in WSSS but also the potential of prompt learning for this problem. The code is available at https://anonymous.4open.science/r/WSS_POLE-DB45/README.md