WACV 2023 Papers

Binary Neural Networks (BNNs) are receiving an upsurge of attention for bringing power-hungry deep learning towards edge devices. The traditional wisdom in this space is to employ sign() for binarizing featuremaps. We argue and illustrate that sign() is a uniqueness bottleneck, limiting information propagation throughout the network. To alleviate this, we propose to dispense sign(), replacing it with a learnable activation binarizer (LAB), allowing the network to learn a fine-grained binarization kernel per layer - as opposed to global thresholding. LAB is a novel universal module that can seamlessly be integrated into existing architectures. To confirm this, we plug it into four seminal BNNs and show a considerable performance boost at the cost of tolerable increase in delay and complexity. Finally, we build an end-to-end BNN (coined as LAB-BNN) around LAB, and demonstrate that it achieves competitive performance on par with the state-of-the-art on ImageNet. Codebase in the supplementary will be made publicly available upon acceptance.

In microscopy image cell segmentation, it is common to train a deep neural network on source data, containing different types of microscopy images, and then fine-tune it using a support set comprising a few randomly selected and annotated training target images. In this paper, we argue that the random selection of unlabelled training target images to be annotated and included in the support set may not enable an effective fine-tuning process, so we propose a new approach to optimise this image selection process. Our approach involves a new scoring function to find informative unlabelled target images. In particular, we propose to measure the consistency in the model predictions on target images against specific data augmentations. However, we observe that the model trained with source datasets does not reliably evaluate consistency on target images. To alleviate this problem, we propose novel self-supervised pretext tasks to compute the scores of unlabelled target images. Finally, the top few images with the least consistency scores are added to the support set for oracle (i.e., expert) annotation and later used to fine-tune the model to the target images. In our evaluations that involve the segmentation of five different types of cell images, we demonstrate promising results on several target test sets compared to the random selection approach as well as other selection approaches, such as Shannon's entropy and Monte-Carlo dropout. Our code will be made publicly available.

Previous video-based human pose estimation methods have shown promising results by leveraging aggregated features of consecutive frames. However, most approaches compromise accuracy to mitigate jitter or do not sufficiently comprehend the temporal aspects of human motion. Furthermore, occlusion increases uncertainty between consecutive frames, which results in unsmooth results. To address these issues, we design an architecture that exploits the keypoint kinematic features with the following components. First, we effectively capture the temporal features by leveraging individual keypoint's velocity and acceleration. Second, the proposed hierarchical transformer encoder aggregates spatio-temporal dependencies and refines the 2D or 3D input pose estimated from existing estimators. Finally, we provide an online cross-supervision between the refined input pose generated from the encoder and the final pose from our decoder to enable joint optimization. We demonstrate comprehensive results and validate the effectiveness of our model in various tasks: 2D pose estimation, 3D pose estimation, body mesh recovery, and sparsely annotated multi-human pose estimation. Our code is available at https://github.com/KyungMinJin/HANet.

Deep image inpainting has made impressive progress with recent advances in image generation and processing algorithms. We claim that the performance of inpainting algorithms can be better judged by the generated structures and textures. Structures refer to the generated object boundary or novel geometric structures within the hole, while texture refers to high-frequency details, especially man-made repeating patterns filled inside the structural regions. We believe that better structures are usually obtained from a coarse-to-fine GAN-based generator network while repeating patterns nowadays can be better modeled using state-of-the-art high-frequency fast fourier convolutional layers. In this paper, we propose a novel inpainting network combining the advantages of the two designs. Therefore, our model achieves a remarkable visual quality to match state-of-the-art performance in both structure generation and repeating texture synthesis using a single network. Extensive experiments demonstrate the effectiveness of the method, and our conclusions further highlight the two critical factors of image inpainting quality, structures, and textures, as the future design directions of inpainting networks.

Burst super-resolution technique provides a possibility of restoring rich details from low-quality images. However, since real world low-resolution (LR) images in practical applications have multiple complicated and unknown degradations, existing non-blind (e.g., bicubic) designed networks usually suffer severe performance drop in recovering high-resolution (HR) images. In this paper, we address the problem of reconstructing HR images from raw burst sequences acquired from modern handheld devices. The central idea is a kernel-guided strategy which can solve the burst SR problem with two steps: kernel estimation and HR image restoration. The former estimates burst kernels from raw inputs, while the latter predicts the super-resolved image based on the estimated kernels. Furthermore, we introduce a pyramid kernel-aware deformable alignment module which can effectively align the raw images with consideration of the blurry priors. Extensive experiments on synthetic and real-world datasets demonstrate that the proposed method can perform favorable state-of-the-art performance in the burst SR problem.

Visual Question Generation (VQG) is a task to generate questions from images. When humans ask questions about an image, their goal is often to acquire some new knowledge. However, existing studies on VQG have mainly addressed question generation from answers or question categories, overlooking the objectives of knowledge acquisition. To introduce a knowledge acquisition perspective into VQG, we constructed a novel knowledge-aware VQG dataset called K-VQG. This is the first large, humanly annotated dataset in which questions regarding images are tied to structured knowledge. We also developed a new VQG model that can encode and use knowledge as the target for a question. The experiment results show that our model outperforms existing models on the K-VQG dataset. Our dataset is publicly available at https://uehara-mech.github.io/kvqg.

A single-shot multispectral camera equipped with an optimized color filter array (CFA) has the potential to deliver a fast and low-cost hyperspectral (HS) imaging system. Previous solutions are largely restricted to designing demosaicing algorithms for fixed CFAs - be it the Bayer color pattern or evenly-spaced spectral multiplexing patterns. Since sampling and reconstruction are tightly-coupled, the ability to search for an optimal solution is severely constrained by using predefined CFAs. In this work, we simultaneously address the problem of spectral band selection, CFA design, image demosaicing, and spectral image recovery in a joint learning framework for single-shot HS imaging. We propose a reinforcement learning (RL) based method for spectral band selection and a novel neural network for CFA generation, image demosaicing, and HS image recovery. The final spectral reconstruction accuracy is used to supervise the training of the main network to maximize the synergies between those tightly-related tasks. The RL method regards the main network as an agent to collect reward. Our final method delivers a simple setup - as simple as an RGB camera - for HS imaging. Experimental results show that our method outperforms competing methods by a large margin.

With the prevalence of CMOS cameras in many computer vision applications, there is an increase in the appearance of rolling shutter (RS) artifacts in captured videos. However, existing video super-resolution algorithms assume that the motion is globally consistent in each video frame and no rolling shutter effect is present. The problem of video super-resolution for video captured using RS cameras is challenging as the model needs to learn the row-wise local pixel displacements and the global structure of the frame for RS correction and super-resolution, simultaneously. Different from existing works, we address a more realistic problem of joint rolling shutter correction and super-resolution (RS-SR). We introduce a novel architecture, deformable Patch Attention Network (PatchNet), that utilizes patch-recurrence property along with deformable receptive fields to learn the global and local structure of the video. Specifically, PatchNet leverages bi-directional motion field in the feature space to extract relevant information from neighboring patches using attention mechanism, and deformable fields using deformable convolutions to extract local pixel-level information for joint rolling shutter correction and super-resolution. Our work is the first to tackle the task of RS correction and super-resolution on the recently released BS-RSCD dataset. Experiments on the BS-RSCD dataset and FastecRS datasets demonstrate that our model performs favorably against various state-of-the-art approaches.

Designing robust algorithms capable of training accurate neural networks on uncurated datasets from the web has been the subject of much research as it reduces the need for time consuming human labor. The focus of many previous research contributions has been on the detection of different types of label noise; however, this paper proposes to improve the correction accuracy of noisy samples once they have been detected. In many state-of-the-art contributions, a two phase approach is adopted where the noisy samples are detected before guessing a corrected pseudo-label in a semi-supervised fashion. The guessed pseudo-labels are then used in the supervised objective without ensuring that the label guess is likely to be correct. This can lead to confirmation bias, which reduces the noise robustness. Here we propose the pseudo-loss, a simple metric that we find to be strongly correlated with pseudo-label correctness on noisy samples. Using the pseudo-loss, we dynamically down weight under-confident pseudo-labels throughout training to avoid confirmation bias and improve the network accuracy. We additionally propose to use a confidence guided contrastive objective that learns robust representation on an interpolated objective between class bound (supervised) for confidently corrected samples and unsupervised representation for under-confident label corrections. Experiments demonstrate the state-of-the-art performance of our Pseudo-Loss Selection (PLS) algorithm on a variety of benchmark datasets including curated data synthetically corrupted with in-distribution and out-of-distribution noise, and two real world web noise datasets. Our experiments are fully reproducible github.com/PaulAlbert31/PLS.

Network architecture plays a pivotal role in the performance of style transfer algorithms. Most existing algorithms use VGG19 as the feature extractor, which incurs a high computational cost. In this work, we conduct an empirical study on the popular network architectures and find that some more efficient networks can replace VGG19 while having comparable style transfer performance. Beyond that, we show that an efficient network can be further accelerated by removing its empty channels via a simple channel pruning method tweaked for style transfer. To prevent the potential performance drop due to using a more lightweight network and obtain better style transfer results, we introduce a more accurate deep feature alignment strategy to improve existing style transfer modules. Taking GoogLeNet as an exemplary efficient network, the pruned GoogLeNet with the improved style transfer module is 2.3 107.4x faster than the state-of-the-art approaches and can achieve 68.03 FPS on 512 x 512 images. Extensive experiments demonstrate that VGG19 can be replaced by a more lightweight network with significantly improved efficiency and comparable style transfer quality.

The generalization with respect to domain shifts, as they frequently appear in applications such as autonomous driving, is one of the remaining big challenges for deep learning models. Therefore, we propose an intra-source style augmentation (ISSA) method to improve domain generalization in semantic segmentation. Our method is based on a novel masked noise encoder for StyleGAN2 inversion. The model learns to faithfully reconstruct the image preserving its semantic layout through noise prediction. Random masking of the estimated noise enables the style mixing capability of our model, i.e. it allows to alter the global appearance without affecting the semantic layout of an image. Using the proposed masked noise encoder to randomize style and content combinations in the training set, ISSA effectively increases the diversity of training data and reduces spurious correlation. As a result, we achieve up to 12.4% mIoU improvements on driving-scene semantic segmentation under different types of data shifts, i.e., changing geographic locations, adverse weather conditions, and day to night. ISSA is model-agnostic and straightforwardly applicable with CNNs and Transformers. It is also complementary to other domain generalization techniques, e.g., it improves the recent state-of-the-art solution RobustNet by 3% mIoU in Cityscapes to Dark Zurich.

Unified panoptic segmentation methods are achieving state-of-the-art results on several datasets. To achieve these results on high-resolution datasets, these methods apply crop-based training. In this work, we find that, although crop-based training is advantageous in general, it also has a harmful side-effect. Specifically, it limits the ability of unified networks to discriminate between large object instances, causing them to make predictions that are confused between multiple instances. To solve this, we propose Intra-Batch Supervision (IBS), which improves a network's ability to discriminate between instances by introducing additional supervision using multiple images from the same batch. We show that, with our IBS, we successfully address the confusion problem and consistently improve the performance of unified networks. For the high-resolution Cityscapes and Mapillary Vistas datasets, we achieve improvements of up to +2.5 on the Panoptic Quality for thing classes, and even more considerable gains of up to +5.8 on both the pixel accuracy and pixel precision, which we identify as better metrics to capture the confusion problem.

Adversarial learning is commonly used to extract latent data representations which are expressive to predict the target attribute but indistinguishable in the privacy attribute. However, whether they can achieve an expected privacy-utility tradeoff is of great uncertainty. In this paper, we posit it is the complex interaction between different attributes in the training set that causes disparate tradeoff results. We first formulate the measurement of utility, privacy and their tradeoff in adversarial learning. Then we propose the metrics of Statistical Reliability (SR) and Feature Reliability (FR) to quantify the relationship between attributes. Specifically, SR reflects the co-occurrence sampling bias of the joint distribution between two attributes. Beyond the explicit dependence, FR exploits the intrinsic interaction one attribute exerts on the other via exploring the representation disentanglement. We validate the metrics in an adversarial learning scheme on CelebA and LFW dataset with a suite of target-privacy attribute pairs. Experiments demonstrate the strong correlations between the metrics and utility, privacy and their tradeoff. We further conclude how to use SR and FR as a guide to the selection of the privacy-utility tradeoff parameter.

We present a new and general framework for convolutional neural network operations on spherical (or omnidirectional) images. Our approach represents the surface as a graph of connected points that doesn't rely on a particular sampling strategy. Additionally, by using an interpolated version of SelectionConv, we can operate on the sphere while using existing 2D CNNs and their weights. Since our method leverages existing graph implementations, it is also fast and can be fine-tuned efficiently. Our method is also general enough to be applied to any surface type, even those that are topologically non-simple. We demonstrate the effectiveness of our technique on the tasks of style transfer and segmentation for spheres as well as stylization for 3D meshes. We provide a thorough ablation study of the performance of various spherical sampling strategies.

Recently, text-guided image manipulation has received increasing attention in the research field of multimedia processing and computer vision due to its high flexibility and controllability. Its goal is to semantically manipulate parts of an input reference image according to the text descriptions. However, most of the existing works have the following problems: (1) text-irrelevant content cannot always be maintained but randomly changed, (2) the performance of image manipulation still needs to be further improved, (3) only can manipulate descriptive attributes. To solve these problems, we propose a novel image manipulation method that interactively edits an image using complex text instructions. It allows users to not only improve the accuracy of image manipulation but also achieve complex tasks such as enlarging, dwindling, or removing objects and replacing the background with the input image. To make these tasks possible, we apply three strategies. First, the given image is divided into text-relevant content and text-irrelevant content. Only the text-relevant content is manipulated and the text-irrelevant content can be maintained. Second, a super-resolution method is used to enlarge the manipulation region to further improve the operability and to help manipulate the object itself. Third, a user interface is introduced for editing the segmentation map interactively to re-modify the generated image according to the user's desires. Extensive experiments on the Caltech-UCSD Birds-200-2011 (CUB) dataset and Microsoft Common Objects in Context (MS COCO) datasets demonstrate our proposed method can enable interactive, flexible, and accurate image manipulation in real-time. Through qualitative and quantitative evaluations, we show that the proposed model outperforms other state-of-the-art methods.

3D hand-object pose estimation is the key to the success of many computer vision applications. The main focus of this task is to effectively model the interaction between the hand and an object. To this end, existing works either rely on interaction constraints in a computationally-expensive iterative optimization, or consider only a sparse correlation between sampled hand and object keypoints. In contrast, we propose a novel dense mutual attention mechanism that is able to model fine-grained dependencies between the hand and the object. Specifically, we first construct the hand and object graphs according to their mesh structures. For each hand node, we aggregate features from every object node by the learned attention and vice versa for each object node. Thanks to such dense mutual attention, our method is able to produce physically plausible poses with high quality and real-time inference speed. Extensive quantitative and qualitative experiments on large benchmark datasets show that our method outperforms state-of-the-art methods. The code is available at https://github.com/rongakowang/DenseMutualAttention.git.

To anticipate how a person would act in the future, it is essential to understand the human intention since it guides the subject towards a certain action. In this paper, we propose a hierarchical architecture which assumes a sequence of human action (low-level) can be driven from the human intention (high-level). Based on this, we deal with long-term action anticipation task in egocentric videos. Our framework first extracts this low- and high-level human information over the observed human actions in a video through a Hierarchical Multi-task Multi-Layer Perceptrons Mixer (H3M). Then, we constrain the uncertainty of the future through an Intention-Conditioned Variational Auto-Encoder (I-CVAE) that generates multiple stable predictions of the next actions that the observed human might perform. By leveraging human intention as high-level information, we claim that our model is able to anticipate more time-consistent actions in the long-term, thus improving the results over the baseline in Ego4D dataset. This work results in the state-of-the-art for Long-Term Anticipation (LTA) task in Ego4D by providing more plausible anticipated sequences, improving the anticipation scores of nouns and actions. Our work ranked first in both CVPR@2022 and ECCV@2022 Ego4D LTA Challenge.

Noisy labels are unavoidable yet troublesome in the ecosystem of deep learning because models can easily overfit them. There are many types of label noise, such as symmetric, asymmetric and instance-dependent noise (IDN), with IDN being the only type that depends on image information. Such dependence on image information makes IDN a critical type of label noise to study, given that labelling mistakes are caused in large part by insufficient or ambiguous information about the visual classes present in images. Aiming to provide an effective technique to address IDN, we present a new graphical modelling approach called InstanceGM, that combines discriminative and generative models. The main contributions of InstanceGM are: i) the use of the continuous Bernoulli distribution to train the generative model, offering significant training advantages, and ii) the exploration of a state-of-the-art noisy-label discriminative classifier to generate clean labels from instance-dependent noisy-label samples. InstanceGM is competitive with current noisy-label learning approaches, particularly in instance-dependent noise benchmarks using synthetic and real-world datasets, where our method shows better accuracy than the competitors in most experiments.

Several recent studies have shown that the use of extra in-distribution data can lead to a high level of adversarial robustness. However, there is no guarantee that it will always be possible to obtain sufficient extra data for a selected dataset. In this paper, we propose a biased multi-domain adversarial training (BiaMAT) method that induces training data amplification on a primary dataset using publicly available auxiliary datasets, without requiring the class distribution match between the primary and auxiliary datasets. The proposed method can achieve increased adversarial robustness on a primary dataset by leveraging auxiliary datasets via multi-domain learning. Specifically, data amplification on both robust and non-robust features can be accomplished through the application of BiaMAT as demonstrated through a theoretical and empirical analysis. Moreover, we demonstrate that while existing methods are vulnerable to negative transfer due to the distributional discrepancy between auxiliary and primary data, the proposed method enables neural networks to flexibly leverage diverse image datasets for adversarial training by successfully handling the domain discrepancy through the application of a confidence-based selection strategy. The code and pre-trained models of our study are available at: https://github.com/BiaMAT/BiaMAT_under_review.