WACV 2022 Papers

Unsupervised anomaly detection with localization has many practical applications when labeling is infeasible and, moreover, when anomaly examples are completely missing in the train data. While recently proposed models for such data setup achieve high accuracy metrics, their complexity is a limiting factor for real-time processing. In this paper, we propose a real-time model and analytically derive its relationship to prior methods. Our CFLOW-AD model is based on a conditional normalizing flow framework adopted for anomaly detection with localization. In particular, CFLOW-AD consists of a discriminatively pretrained encoder followed by a multi-scale generative decoders where the latter explicitly estimate likelihood of the encoded features. Our approach results in a computationally and memory-efficient model: CFLOW-AD is faster and smaller by a factor of 10x than prior state-of-the-art with the same input setting. Our experiments on the MVTec dataset show that CFLOW-AD outperforms previous methods by 0.36% AUROC in detection task, by 1.12% AUROC and 2.5% AUPRO in localization task, respectively. We open-source our code with fully reproducible experiments.

Image-based virtual try-on techniques have shown great promise for enhancing the user-experience and improving customer satisfaction on fashion-oriented e-commerce platforms. However, they are currently still limited in the quality of the try-on results they are able to produce from input images of diverse characteristics. In this work, we propose a Context-Driven Virtual Try-On Network (C-VTON) that addresses these limitations and convincingly transfers selected clothing items to the target subjects even under challenging pose configurations and in the presence of self-occlusions. At the core of the C-VTON pipeline are: (i) a geometric matching procedure that efficiently aligns the target clothing with the pose of the person in the input images, and (ii) a powerful image generator that utilizes various types of contextual information when synthesizing the final try-on result. C-VTON is evaluated in rigorous experiments on the VITON and MPV datasets and in comparison to state-of-the-art techniques from the literature. Experimental results show that the proposed approach is able to produce photo-realistic and visually convincing results and significantly improves on the existing state-of-the-art.

In video data, busy motion details from moving regions are conveyed within a specific frequency bandwidth in the frequency domain. Meanwhile, the rest of the frequencies of video data are encoded with quiet information with substantial redundancy, which causes low processing efficiency in existing video models that take as input raw RGB frames. In this paper, we consider allocating intenser computation for the processing of the important busy information and less computation for that of the quiet information. We design a trainable Motion Band-Pass Module (MBPM) for separating busy information from quiet information in raw video data. By embedding the MBPM into a two-pathway CNN architecture, we define a Busy-Quiet Net (BQN). The efficiency of BQN is determined by avoiding redundancy in the feature space processed by the two pathways: one operating on Quiet features of low-resolution, while the other processes Busy features. The proposed BQN outperforms many recent video processing models on Something-Something V1, Kinetics400, UCF101 and HMDB51 datasets.

Self-supervised representation learning has made significant leaps fueled by progress in contrastive learning, which seeks to learn transformations that embed positive input pairs nearby, while pushing negative pairs far apart. While positive pairs can be generated reliably (e.g., as different views of the same image), it is difficult to accurately establish negative pairs, defined as samples from different images regardless of their semantic content or visual features. A fundamental problem in contrastive learning is mitigating the effects of false negatives. Contrasting false negatives induces two critical issues in representation learning: discarding semantic information and slow convergence. In this paper, we propose novel approaches to identify false negatives, as well as two strategies to mitigate their effect, i.e. false negative elimination and attraction, while systematically performing rigorous evaluations to study this problem in detail. Our method exhibits consistent improvements over existing contrastive learning-based methods. Without labels, we identify false negatives with 40% accuracy among 1000 semantic classes on ImageNet, and achieve 5.8% absolute improvement in top-1 accuracy over the previous state-of-the-art when finetuning with 1% labels.

With population growth and a shrinking rural workforce, agricultural technologies have become increasingly important. Above-ground biomass (AGB) is a key trait relevant to breeding, agronomy and crop physiology field experiments. However, measuring the biomass of a cereal plot requires cutting, drying and weighing processes, which are laborious, expensive and destructive tasks. This paper proposes a non-destructive and high-throughput method to predict biomass from field samples based on Light Detection and Ranging (LiDAR). Unlike previous methods that are based on the density of a point cloud or plant height, our biomass prediction network (BioNet) additionally considers plant structure. Our BioNet contains three modules: 1) a completion module to predict missing points due to canopy occlusion; 2) a regularization module to regularize the neural representation of the whole plot; and 3) a projection module to learn the salient structures from a bird's eye view of the point cloud. An attention-based fusion block is used to achieve final biomass predictions. In addition, the complete dataset, including hand-measured biomass and LiDAR data, is made available to the community. Experiments show that our BioNet achieves approximately 33% improvement over current state-of-the-art methods.

Large-scale pretraining of visual representations has led to state-of-the-art performance on a range of benchmark computer vision tasks, yet the benefits of these techniques at extreme scale in complex production systems has been relatively unexplored. We consider the case of a popular visual discovery product, where these representations are trained with multi-task learning, from use-case specific visual understanding (e.g. skin tone classification) to general representation learning for all visual content (e.g. embeddings for retrieval). In this work, we describe how we (1) generate a dataset with over a billion images via large weakly-supervised pretraining to improve the performance of these visual representations, and (2) leverage Transformers to replace the traditional convolutional backbone, with insights into both system and performance improvements, especially at 1B+ image scale. To support this backbone model, we detail a systematic approach to deriving weakly-supervised image annotations from heterogenous text signals, demonstrating the benefits of clustering techniques to handle the long-tail distribution of image labels. Through a comprehensive study of offline and online evaluation, we show that large-scale Transformer-based pretraining provides significant benefits to industry computer vision applications. The model is deployed in a production visual shopping system, with 36% improvement in top-1 relevance and 23% improvement in click-through volume. We conduct extensive experiments to better understand the empirical relationships between Transformer-based architectures, dataset scale, and the performance of production vision systems.

The advancement in numerous generative models has a two-fold effect: a simple and easy generation of realistic synthesized images, but also an increased risk of malicious abuse of those images. Thus, it is important to develop a generalized detector for synthesized images of any GAN model or object category, including those unseen during the training phase. However, the conventional methods heavily depend on the training settings, which cause a dramatic decline in performance when tested with unknown domains. To resolve the issue and obtain a generalized detection ability, we propose Bilateral High-Pass Filters (BiHPF), which amplify the effect of the frequency-level artifacts that are generally found in the synthesized images of generative models. Also, to find the properties of the general frequency-level artifacts, we develop an additional method to adversarially extract the artifact compression map. Numerous experimental results validate that our method outperforms other state-of-the-art methods, even when tested with unseen domains.

Binaural audio gives the listener an immersive experience and can enhance augmented and virtual reality. However, recording binaural audio requires specialized setup with a dummy human head having microphones in left and right ears. Such a recording setup is difficult to build and setup, therefore mono audio has become the preferred choice in common devices. To obtain the same impact as binaural audio, recent efforts have been directed towards lifting mono audio to binaural audio conditioned on the visual input from the scene. Such approaches have not used an important cue for the task: the distance of different sound producing objects from the microphones. In this work, we argue that depth map of the scene can act as a proxy for inducing distance information of different objects in the scene, for the task of audio binauralization. We propose a novel encoder-decoder architecture with a hierarchical attention mechanism to encode image, depth and audio feature jointly. We design the network on top of state-of-the-art transformer networks for image and depth representation. We show empirically that the proposed method outperforms state-of-the-art methods comfortably for two challenging public datasets FAIR-Play and MUSIC- Stereo. We also demonstrate with qualitative results that the method is able to focus on the right information required for the task. The qualitative results are available at our project page https://krantiparida.github.io/projects/bmonobinaural.html

Active learning algorithms select a subset of data for annotation to maximize the model performance on a budget. One such algorithm is Expected Gradient Length, which as the name suggests uses the approximate gradient induced per example in the sampling process. While Expected Gradient Length has been successfully used for classification and regression, the formulation for regression remains intuitively driven. Hence, our theoretical contribution involves deriving this formulation, thereby supporting experimental evidence [4, 5]. Subsequently, we show that expected gradient length in regression is equivalent to Bayesian uncertainty [22]. If certain assumptions are infeasible, our algorithmic contribution (EGL++) approximates the effect of ensembles with a single deterministic network. Instead of computing multiple possible inferences per input, we leverage previously annotated samples to quantify the probability of previous labels being the true label. Such an approach allows us to extend expected gradient length to a new task: human pose estimation. We perform experimental validation on two human pose datasets (MPII and LSP/LSPET), highlighting the interpretability and competitiveness of EGL++ with different active learning algorithms for human pose estimation.

The goal of filter pruning is to search for unimportant filters to remove in order to make convolutional neural networks (CNNs) efficient without sacrificing the performance in the process. The challenge lies in finding information that can help determine how important or relevant each filter is with respect to the final output of neural networks. In this work, we share our observation that the batch normalization (BN) parameters of pre-trained CNNs can be used to estimate the feature distribution of activation outputs, without processing of training data. Upon observation, we propose a simple yet effective filter pruning method by evaluating the importance of each filter based on the BN parameters of pre-trained CNNs. The experimental results on CIFAR-10 and ImageNet demonstrate that the proposed method can achieve outstanding performance with and without fine-tuning in terms of the trade-off between the accuracy drop and the reduction in computational complexity and number of parameters of pruned networks.

In video segmentation, generating temporally consistent results across frames is as important as achieving frame-wise accuracy. Existing methods rely either on optical flow regularization or fine-tuning with test data to attain temporal consistency. However, optical flow is not always avail-able and reliable. Besides, it is expensive to compute. Fine-tuning the original model in test time is cost sensitive. This paper presents an efficient, intuitive, and unsupervised online adaptation method, AuxAdapt, for improving the temporal consistency of most neural network models. It does not require optical flow and only takes one pass of the video. Since inconsistency mainly arises from the model's uncertainty in its output, we propose an adaptation scheme where the model learns from its own segmentation decisions as it streams a video, which allows producing more confident and temporally consistent labeling for similarly-looking pixels across frames. For stability and efficiency, we leverage a small auxiliary segmentation network (AuxNet) to assist with this adaptation. More specifically, AuxNet readjusts the decision of the original segmentation network (Main-Net) by adding its own estimations to that of MainNet. At every frame, only AuxNet is updated via back-propagation while keeping MainNet fixed. We extensively evaluate our test-time adaptation approach on standard video benchmarks, including Cityscapes, CamVid, and KITTI. The results demonstrate that our approach provides label-wise accurate, temporally consistent, and computationally efficient adaptation (5+ folds overhead reduction comparing to state-of-the-art test-time adaptation methods).

In the semiconductor industry, reverse engineering is used to extract information from microchips. Circuit extraction is becoming increasingly difficult due to the continuous technology shrinking. A high quality reverse engineering process is challenged by various defects coming from chip preparation and imaging errors. Currently, no automated, technology-agnostic defect inspection framework is available. To meet the requirements of the mostly manual reverse engineering process, the proposed automated framework needs to handle highly imbalanced data, as well as unknown and multiple defect classes. We propose a network architecture that is composed of a shared Xception-based feature extractor and multiple, individually trainable binary classification heads: the HydREnet. We evaluated our defect classifier on three challenging industrial datasets and achieved accuracies of over 85 %, even for underrepresented classes. With this framework, the manual inspection effort can be reduced down to 5 %.

Efficient video architecture is the key to the deployment of video action recognition systems on devices with limited computing capabilities. Unfortunately, existing video architectures are often computationally intensive and not suitable for such applications. The recent X3D work presents a new family of efficient video models by expanding a hand-crafted image architecture along multiple axes, such as space, time, width, and depth. Although operating in a conceptually large space, X3D searched one axis at a time, and merely explored a small set of 30 architectures in total, which does not sufficiently explore the space. This paper bypasses existing 2D architectures, and directly searched for 3D architectures in a fine-grained space, where block type, filter number, expansion ratio and attention block are jointly searched. A probabilistic neural architecture search method is adopted to efficiently search in such a large space. Evaluations on Kinetics and Something-Something-V2 benchmarks confirm our \autoxthreed models outperform existing ones in accuracy up to 1.7% under similar FLOPs, and reduce the computational cost up to 1.74 times to reach similar performance. Code will be publicly available.

Auto white balance (AWB) is applied by camera hardware at capture time to remove the color cast caused by the scene illumination. The vast majority of white-balance algorithms assume a single light source illuminates the scene; however, real scenes often have mixed lighting conditions. This paper presents an effective AWB method to deal with such mixed-illuminant scenes. A unique departure from conventional AWB, our method does not require illuminant estimation, as is the case in traditional camera AWB modules. Instead, our method proposes to render the captured scene with a small set of predefined white-balance settings. Given this set of rendered images, our method learns to estimate weighting maps that are used to blend the rendered images to generate the final corrected image. Through extensive experiments, we show this proposed method produces promising results compared to other alternatives for single- and mixed-illuminant scene color correction.

In this paper, we audit saliency cropping algorithms used by Twitter, Google and Apple to investigate issues pertaining to the male-gaze cropping phenomenon as well as race-gender biases that emerge in post-cropping survival ratios of face-images constituting 3 x 1 grid images. In doing so, we present the first formal empirical study which suggests that the worry of a male-gaze-like image cropping phenomenon on Twitter is not at all far-fetched and it does occur with worryingly high prevalence rates in real-world full-body single-female-subject images shot with logo-littered backdrops. We uncover that while all three saliency cropping frameworks considered in this paper do exhibit acute racial and gender biases, Twitter's saliency cropping framework uniquely elicits high male-gaze cropping prevalence rates. In order to facilitate reproducing the results presented here, we are open-sourcing both the code and the datasets that we curated at shorturl.at/iuzK9. We hope the computer vision community and saliency cropping researchers will build on the results presented here and extend these investigations to similar frameworks deployed in the real world by other companies such as Microsoft and Facebook.

In recent years, periocular recognition has been developed as a valuable biometric identification approach, especially in wild environments (for example, masked faces due to COVID-19 pandemic) where facial recognition may not be applicable. This paper presents a new deep periocular recognition framework called attribute-based deep periocular recognition (ADPR), which predicts soft biometrics and incorporates the prediction into a periocular recognition algorithm to determine identity from periocular images with high accuracy. We propose an end-to-end framework, which uses several shared convolutional neural network (CNN) layers (a common network) whose output feeds two separate dedicated branches (modality dedicated layers); the first branch classifies periocular images while the second branch predicts soft biometrics. Next, the features from these two branches are fused together for a final periocular recognition. The proposed method is different from existing methods as it not only uses a shared CNN feature space to train these two tasks jointly, but it also fuses predicted soft biometric features with the periocular features in the training step to improve the overall periocular recognition performance. Our proposed model is extensively evaluated using four different publicly available datasets. Experimental results indicate that our soft biometric based periocular recognition approach outperforms other state-of-the-art methods for periocular recognition in wild environments.

Whilst adversarial attack detection has received considerable attention, it remains a fundamentally challenging problem from two perspectives. First, while threat models can be well-defined, attacker strategies may still vary widely within those constraints. Therefore, detection should be considered as an open-set problem, standing in contrast to most current detection approaches. These methods take a closed-set view and train binary detectors, thus biasing detection toward attacks seen during detector training. Second, limited information is available at test time and typically confounded by nuisance factors including the label and underlying content of the image. We address these challenges via a novel strategy based on random subspace analysis. We present a technique that utilizes properties of random projections to characterize the behavior of clean and adversarial examples across a diverse set of subspaces. The self-consistency (or inconsistency) of model activations is leveraged to discern clean from adversarial examples. Performance evaluations demonstrate that our technique (AUC [0.92, 0.98]) outperforms competing detection strategies (AUC [0.30,0.79]), while remaining truly agnostic to the attack strategy (for both targeted/untargeted attacks). It also requires significantly less calibration data (composed only of clean examples) than competing approaches to achieve this performance.

Mesh representation by random walks has been shown to benefit deep learning. Randomness is indeed a powerful concept. However, it comes with a price--some walks might wander around non-characteristic regions of the mesh, which might be harmful to shape analysis, especially when only a few walks are utilized. We propose a novel walk-attention mechanism that leverages the fact that multiple walks are used for a single mesh representation. The key idea is that the walks may provide each other with information regarding the meaningful (attentive) features of the mesh. We utilize this mutual information to extract a single descriptor of the mesh. This differs from common attention mechanisms that use attention to improve the representation of each individual descriptor. Our approach achieves SoTA results for two basic 3D shape analysis tasks: classification and retrieval. Even a handful of walks along a mesh suffice for learning. Furthermore, our approach provides insight into mesh importance detection.

The standard paradigm in neural architecture search (NAS) is to search for a fully deterministic architecture with specific operations and connections. In this work, we instead propose to search for the optimal operation distribution, thus providing a stochastic and approximate solution, which can be used to sample architectures of arbitrary length. We propose and show, that given an architectural cell, its performance largely depends on the ratio of used operations, rather than any specific connection pattern; that is, small changes in the ordering of the operations are often irrelevant. This intuition is orthogonal to any specific search strategy and can be applied to a diverse set of NAS algorithms. Through extensive validation on 4 data-sets and 4 NAS techniques (Bayesian optimisation, differentiable search, local search and random search), we show that the operation distribution (1) holds enough discriminating power to reliably identify a solution and (2) is significantly easier to optimise than traditional encodings, leading to large speed-ups at little to no cost in performance. Indeed, this simple intuition significantly reduces the cost of current approaches and potentially enable NAS to be used in a broader range of research applications.