WACV 2024 Papers

Weakly supervised semantic segmentation (WSSS) employing weak forms of labels has been actively studied to alleviate the annotation cost of acquiring pixel-level labels. However, classifiers trained on biased datasets tend to exploit shortcut features and make predictions based on spurious correlations between certain backgrounds and objects, leading to a poor generalization performance. In this paper, we propose shortcut mitigating augmentation (SMA) for WSSS, which generates synthetic representations of object-background combinations not seen in the training data to reduce the use of shortcut features. Our approach disentangles the object-relevant and background features. We then shuffle and combine the disentangled representations to create synthetic features of diverse object-background combinations. SMA-trained classifier depends less on contexts and focuses more on the target object when making predictions. In addition, we analyzed the behavior of the classifier on shortcut usage after applying our augmentation using an attribution method-based metric. The proposed method achieved the improved performance of semantic segmentation result on PASCAL VOC 2012 and MS COCO 2014 datasets.

Video Object Segmentation (VOS) is crucial for several applications, from video editing to video data generation. Training a VOS model requires an abundance of manually labeled training videos. The de-facto traditional way of annotating objects requires humans to draw detailed segmentation masks on the target objects at each video frame. This annotation process, however, is tedious and time-consuming. To reduce this annotation cost, in this paper, we propose EVA-VOS, a human-in-the-loop annotation framework for video object segmentation. Unlike the traditional approach, we introduce an agent that predicts iteratively both which frame ("What") to annotate and which annotation type ("How") to use. Then, the annotator annotates only the selected frame that is used to update a VOS module, leading to significant gains in annotation time. We conduct experiments on the MOSE and the DAVIS datasets and we show that: (a) EVA-VOS leads to masks with accuracy close to the human agreement 3.5x faster than the standard way of annotating videos; (b) our frame selection achieves state-of-the-art performance; (c) EVA-VOS yields significant performance gains in terms of annotation time compared to all other methods and baselines.

Body shape is a crucial factor in outfit recommendation. Previous studies that directly used body measurement data to investigate the relationship between body shape and outfit have achieved limited performance due to oversimplified body shape representations. This paper proposes a Visual Body-shape-Aware Network (ViBA-Net) to improve the fashion compatibility model's awareness of human body shape through visual-level information. Specifically, ViBA-Net consists of three modules: a body-shape embedding module, which extracts visual and anthropometric features of body shape from a newly introduced large-scale body shape dataset; an outfit embedding module, which learns the outfit representation based on visual features extracted from a try-on image and textual features extracted from fashion attributes; and a joint embedding module, which jointly models the relationship between the representations of body shape and outfit. ViBA-Net is designed to generate attribute-level explanations for the evaluation results based on the computed attention weights. The effectiveness of ViBA-Net is evaluated on two mainstream datasets through qualitative and quantitative analysis. Data and code are released.

Image anomaly localization (IAL) is widely applied in fault detection and industrial inspection domains to discover anomalous patterns in images at the pixel level. The unique challenge of this task is the lack of comprehensive anomaly samples for model training. The state-of-the-art methods train end-to-end models that leverage outlier exposure to simulate pseudo anomalies, but they show poor transferability to new datasets due to the inherent bias to the synthesized outliers during training. Recently, two-stage instance-level self-supervised learning (SSL) has shown potential in learning generic representations for IAL. However, we hypothesize that dense-level SSL is more compatible as IAL requires pixel-level prediction. In this paper, we bridge these gaps by proposing a two-stage, dense pre-training model tailored for the IAL task. More specifically, our model utilizes dual positive-pair selection criteria and dual feature scales to learn more effective representations. Through extensive experiments, we show that our learned representations achieve significantly better anomaly localization performance among two-stage models, while requiring almost half the convergence time. Moreover, our learned representations have better transferability to unseen datasets. Code is available at https://github. com/terrlo/DS2.

Recognizing multiple objects in an image is challenging due to occlusions, and becomes even more so when the objects are small. While promising, existing multi-label image recognition models do not explicitly learn context-based representations, and hence struggle to correctly recognize small and occluded objects. Intuitively, recognizing occluded objects requires knowledge of partial input, and hence context. Motivated by this intuition, we propose Masked Supervised Learning (MSL), a single-stage, model-agnostic learning paradigm for multi-label image recognition. The key idea is to learn context-based representations using a masked branch and to model label co-occurrence using label consistency. Experimental results demonstrate the simplicity, applicability and more importantly the competitive performance of MSL against previous state-of-the-art methods on standard multi-label image recognition benchmarks. In addition, we show that MSL is robust to random masking and demonstrate its effectiveness in recognizing non-masked objects. Code and pretrained models are available on GitHub.

Semantic segmentation methods have advanced significantly. Still, their robustness to real-world perturbations and object types not seen during training remains a challenge, particularly in safety-critical applications. We propose a novel approach to improve the robustness of semantic segmentation techniques by leveraging the synergy between label-to-image generators and image-to-label segmentation models. Specifically, we design Robusta, a novel robust conditional generative adversarial network to generate realistic and plausible perturbed images that can be used to train reliable segmentation models. We conduct in-depth studies of the proposed generative model, assess the performance and robustness of the downstream segmentation network, and demonstrate that our approach can significantly enhance the robustness in the face of real-world perturbations, distribution shifts, and out-of-distribution samples. Our results suggest that this approach could be valuable in safety-critical applications, where the reliability of perception modules such as semantic segmentation is of utmost importance and comes with a limited computational budget in inference. We release our code at https://github.com/ENSTA-U2IS/robusta.

Neural parameter allocation search (NPAS) automates parameter sharing by obtaining weights for a network given an arbitrary, fixed parameter budget. Prior work has two major drawbacks we aim to address. First, there is a dis- connect in the sharing pattern between the search and train- ing steps, where weights are warped for layers of different sizes during the search to measure similarity, but not during training, resulting in reduced performance. To address this, we generate layer weights by learning to compose sets of SuperWeights, which represent a group of trainable parameters. These SuperWeights are created to be large enough so they can be used to represent any layer in the network, but small enough that they are computationally efficient. The second drawback we address is the method of measuring similarity between shared parameters. Whereas prior work compared the weights themselves, we argue this does not take into account the amount of conflict between the shared weights. Instead, we use gradient information to identify layers with shared weights that wish to diverge from each other. We demonstrate that our SuperWeight Networks consistently boost performance over the state-of-the-art on the ImageNet and CIFAR datasets in the NPAS setting. We further show that our approach can generate parameters for many network architectures using the same set of weights. This enables us to support tasks like efficient ensembling and anytime prediction, outperforming fully-parameterized ensembles with 17% fewer parameters.

Pre-trained Visual-Language Models (VLMs), such as CLIP, have shown enhanced performance across a range of tasks that involve the integration of visual and linguistic elements. When CLIP is used for depth estimation tasks, the patches, divided from the input images, can be combined with a series of semantic descriptions of the depth information to obtain similarity results. The coarse estimation of depth is then achieved by weighting and summing the depth values, called depth bins, corresponding to the predefined semantic descriptions. The zero-shot approach circumvents the computational and time-intensive nature of traditional fully-supervised depth estimation methods. However, this method, utilizing fixed depth bins, may not effectively generalize as images from different scenes may exhibit distinct depth distributions. To address this challenge, we propose a few-shot-based method which learns to adapt the VLMs for monocular depth estimation to balance training costs and generalization capabilities. Specifically, it assigns different depth bins for different scenes, which can be selected by the model during inference. Additionally, we incorporate learnable prompts to preprocess the input text to convert the easily human-understood text into easily model-understood vectors and further enhance the performance. With only one image per scene for training, our extensive experiment results on the NYU V2 dataset demonstrate that our method outperforms the previous state-of-the-art method by up to 10.6% in terms of MARE.

We present a novel approach for saliency prediction in images, leveraging parallel decoding in transformers to learn saliency solely from fixation maps. Models typically rely on continuous saliency maps, to overcome the difficulty of optimizing for the discrete fixation map. We attempt to replicate the experimental setup that generates saliency datasets. Our approach treats saliency prediction as a direct set prediction problem, via a global loss that enforces unique fixations prediction through bipartite matching and a transformer encoder-decoder architecture. By utilizing a fixed set of learned fixation queries, the cross-attention reasons over the image features to directly output the fixation points, distinguishing it from other modern saliency predictors. Our approach, named Saliency TRansformer (SalTR) achieves remarkable results on the Salicon benchmark.

Decoding the human brain has been a hallmark of neuroscientists and Artificial Intelligence researchers alike. Reconstruction of visual images from brain Electroencephalography (EEG) signals has garnered a lot of interest due to its applications in brain-computer interfacing. This study proposes a two-stage method where the first step is to obtain EEG-derived features for robust learning of deep representations and subsequently utilize the learned representation for image generation and classification. We demonstrate the generalizability of our feature extraction pipeline across three different datasets using deep-learning architectures with supervised and contrastive learning methods. We have performed the zero-shot EEG classification task to support the generalizability claim further. We observed that a subject invariant linearly separable visual representation was learned using EEG data alone in an unimodal setting that gives better k-means accuracy as compared to a joint representation learning between EEG and images. Finally, we propose a novel framework to transform unseen images into the EEG space and reconstruct them with approximation, showcasing the potential for image reconstruction from EEG signals. Our proposed image synthesis method from EEG shows 62.9% and 36.13% inception score improvement on the EEGCVPR40 and the Thoughtviz datasets, which is better than state-of-the-art performance in GAN.

Trendy suggestions for learning-based elastic warps enable the deep image stitchings to align images exposed to large parallax errors. Despite the remarkable alignments, the methods struggle with occasional holes or discontinuity between overlapping and non-overlapping regions of a target image as the applied training strategy mostly focuses on overlap region alignment. As a result, they require additional modules such as seam finder and image inpainting for hiding discontinuity and filling holes, respectively. In this work, we suggest Recurrent Elastic Warps (REwarp) that address the problem with Dirichlet boundary condition and boost performances by residual learning for recurrent misalign correction. Specifically, REwarp predicts a homography and a Thin-plate Spline (TPS) under the boundary constraint for discontinuity and hole-free image stitching. Our experiments show the favorable aligns and the competitive computational costs of REwarp compared to the existing stitching methods. Our source code is available at https://github.com/minshu-kim/REwarp.

Current deep learning paradigms largely benefit from the tremendous amount of annotated data. However, the quality of the annotations often varies among labelers. Multi-observer studies have been conducted to examine the annotation variances (by labeling the same data multiple times) and their effects on critical applications like medical image analysis. In this paper, we demonstrate how multiple sets of annotations (either hand-labeled or algorithm-generated) can be utilized together and mutually benefit the learning of classification tasks. The concept of learning-to-vote is introduced to sample quality label sets for each data entry on-the-fly during the training. Specifically, a meta-training-based label-sampling module is designed to achieve refined labels (weighted sum of attended ones) that benefit the model learning the most through additional back-propagations. We apply the learning-to-vote scheme on the classification task of a synthetic noisy CIFAR-10 to prove the concept and then demonstrate superior results (3-5% increase on average in multiple disease classification AUCs) on the chest x-ray images from a hospital-scale dataset (MIMIC-CXR) and hand-labeled dataset (OpenI) in comparison to regular training paradigms.

The autoencoder is an unsupervised learning paradigm that aims to create a compact latent representation of data by minimizing the reconstruction loss. However, it tends to overlook the fact that most data (images) are embedded in a lower-dimensional latent space, which is crucial for effective data representation. To address this limitation, we propose a novel approach called Low-Rank Autoencoder (LoRAE). In LoRAE, we incorporated a low-rank regularizer to adaptively learn a low-dimensional latent space while preserving the basic objective of an autoencoder. This helps embed the data in a lower-dimensional latent space while preserving important information. It is a simple autoencoder extension that learns low-rank latent space. Theoretically, we establish a tighter error bound for our model. Empirically, our model's superiority shines through various tasks such as image generation and downstream classification. Both theoretical and practical outcomes highlight the importance of acquiring low-dimensional embeddings.

In this paper, we address the challenges of representing feature distributions which have multimodality within a class in deep neural networks. Existing online clustering methods employ sub-centroids to capture intra-class variations. However, conducting online clustering faces some limitations, i.e., online clustering assigns only a single subcentroid to a feature vector extracted from a backbone and ignores the relationship between the other sub-centroids and the feature vector, and updating sub-centroids in an online clustering manner incurs significant storage costs. To address these limitations, we propose a novel method utilizing orthonormal matrices instead of sub-centroids for relaxing discrete assignments into continuous assignments. We update the orthonormal matrices using a gradient-based method, which eliminates the need for online clustering or additional storage. Experimental results on the CIFAR and ImageNet datasets exhibit that the proposed method outperforms current online clustering techniques in classification accuracy, sub-category discovery, and transferability, providing an efficient solution to the challenges posed by complex recognition targets.

No-reference (NR) image quality assessment (IQA) is an important tool in enhancing the user experience in diverse visual applications. A major drawback of state-of-the-art NR-IQA techniques is their reliance on a large number of human annotations to train models for a target IQA application. To mitigate this requirement, there is a need for unsupervised learning of generalizable quality representations that capture diverse distortions. We enable the learning of low-level quality features agnostic to distortion types by introducing a novel quality-aware contrastive loss. Further, we leverage the generalizability of vision-language models by fine-tuning one such model to extract high-level image quality information through relevant text prompts. The two sets of features are combined to effectively predict quality by training a simple regressor with very few samples on a target dataset. Additionally, we design zero-shot quality predictions from both pathways in a completely blind setting. Our experiments on diverse datasets encompassing various distortions show the generalizability of the features and their superior performance in the data-efficient and zero-shot settings.

Single-source open-domain generalization (SS-ODG) addresses the challenge of labeled source domains with supervision during training and unlabeled novel target domains during testing. The target domain includes both known classes from the source domain and samples from previously unseen classes. Existing techniques for SS-ODG primarily focus on calibrating source-domain classifiers to identify open samples in the target domain. However, these methods struggle with visually fine-grained open-closed data, often misclassifying open samples as closed-set classes. Moreover, relying solely on a single source domain restricts the model's ability to generalize. To overcome these limitations, we propose a novel framework called SODG-NET that simultaneously synthesizes novel domains and generates pseudo-open samples using a learning-based objective, in contrast to the ad-hoc mixing strategies commonly found in the literature. Our approach enhances generalization by diversifying the styles of known class samples using a novel metric criterion and generates diverse pseudo-open samples to train a unified and confident multiclass classifier capable of handling both open and closed-set data. Extensive experimental evaluations conducted on multiple benchmarks consistently demonstrate the superior performance of SODG-NET compared to the literature.

We address the problem of keypoint selection, and find that the performance of 6DoF pose estimation methods can be improved when pre-defined keypoint locations are learned, rather than being heuristically selected as has been the standard approach. We found that accuracy and efficiency can be improved by training a graph network to select a set of disperse keypoints with similarly distributed votes. These votes, learned by a regression network to accumulate evidence for the keypoint locations, can be regressed more accurately compared to previous heuristic keypoint algorithms. The proposed KeyGNet, supervised by a combined loss measuring both Wasserstein distance and dispersion, learns the color and geometry features of the target objects to estimate optimal keypoint locations. Experiments demonstrate the keypoints selected by KeyGNet improved the accuracy for all evaluation metrics of all seven datasets tested, for three keypoint voting methods. The challenging Occlusion LINEMOD dataset notably improved ADD(S) by +16.4% on PVN3D, and all core BOP datasets showed an AR improvement for all objects, of between +1% and +21.5%. There was also a notable increase in performance when transitioning from single object to multiple object training using KeyGNet keypoints, essentially eliminating the SISO-MIMO gap for Occlusion LINEMOD.

With the development of cloud services and machine learning, there has been an inevitable need to enhance privacy and security when serving video recognition models. Although existing image encryption methods can be used to address this issue, applying them frame by frame to videos is insufficient in two respects: model performance degradation and security strength. In this paper, we propose a novel encryption approach for privacy-preserving action recognition. It consists of two encrypting operations; Learnable Cube-based Video Encryption (LCVE) and ViT Scrambling. LCVE is video encryption based on spatio-temporal cubes, which has a large key space and can provide robust privacy protection. ViT Scrambling encrypts the Vision Transformer (ViT) model, which enables it to recognize the encrypted videos in the same manner as unencrypted videos without modifying the model architecture or fine-tuning on the encrypted data. We evaluate our method in an action recognition task with seven datasets containing a variety of action classes as well as motion and visual patterns. Empirical results demonstrate that LCVE combined with ViT Scrambling can preserve video privacy while recognizing action in encrypted videos as well as unencrypted videos. As a result, our approach outperforms existing privacy-preserving action recognition methods.

Recent data-privacy laws have sparked interest in machine unlearning, which involves removing the effect of specific training samples from a learnt model as if they were never present in the original training dataset. The challenge of machine unlearning is to discard information about the "forget" data in the learnt model without altering the knowledge about the remaining dataset and to do so more efficiently than the naive retraining approach. To achieve this, we adopt a projected-gradient based learning method, named as Projected-Gradient Unlearning (PGU), in which the model takes steps in the orthogonal direction to the gradient subspaces deemed unimportant for the retaining dataset, so as to its knowledge is preserved. By utilizing Stochastic Gradient Descent (SGD) to update the model weights, our method can efficiently scale to any model and dataset size. We provide empirically evidence to demonstrate that our unlearning method can produce models that behave similar to models retrained from scratch across various metrics even when the training dataset is no longer accessible. Our code is available at https://github.com/hnanhtuan/projected_gradient_unlearning.