WACV 2022 Papers

Image outpainting has recently received considerable attention because it can be useful in tasks such as image retargeting and panorama image generation. In general, the problem of extending an image beyond its given boundaries is still ill-posed. Conventional methods predominantly attempt image outpainting by using complex network structures. Some recent studies have tried to decrease the problem complexity through the conversion techniques from outpainting to inpainting. Although these methodologies work well in simple cases, their performance reduces considerably for asymmetrical images. This paper proposes a novel hint-based outpainting methodology that can adaptively select the most plausible patches as hints from a given image to reduce the difficulty of outpainting. To estimate high-quality hints, inspired by patch-based image inpainting methods, we utilize Vision Transformer that also considers self-attention for each patch. The estimated hints are attached on both boundaries of the input image and the inside missing regions are predicted by using an inpainting network. After finishing the prediction, the output image is obtained by removing the hints. Experiments show that our image-adaptive hint framework, when employed in representative inpainting networks, can consistently improve its performance compared to the other conversion techniques from outpainting to inpainting on SUN and Beach benchmark datasets.

Ill-posed inverse problems appear in many image processing applications, such as deblurring and super-resolution. In recent years, solutions that are based on deep Convolutional Neural Networks (CNNs) have shown great promise. Yet, most of these techniques, which train CNNs using external data, are restricted to the observation models that have been used in the training phase. A recent alternative that does not have this drawback relies on learning the target image using internal learning. One such prominent example is the Deep Image Prior (DIP) technique that trains a network directly on the input image with the least-squares loss. In this paper, we propose a new image restoration framework that is based on minimizing a loss function that includes a "projected-version" of the Generalized Stein Unbiased Risk Estimator (GSURE) and parameterization of the latent image by a CNN. We demonstrate two ways to use our framework. In the first one, where no explicit prior is used, we show that the proposed approach outperforms other internal learning methods, such as DIP. In the second one, we show that our GSURE-based loss leads to improved performance when used within a plug-and-play priors scheme.

In this paper, we introduce the task of multi-view RGB-based 3D object detection as an end-to-end optimization problem. To address this problem, we propose ImVoxelNet, a novel fully convolutional method of 3D object detection based on posed monocular or multi-view RGB images. The number of monocular images in each multi-view input can variate during training and inference; actually, this number might be unique for each multi-view input. ImVoxelNet successfully handles both indoor and outdoor scenes, which makes it general-purpose. Specifically, it achieves state-of-the-art results in car detection on KITTI (monocular) and nuScenes (multi-view) benchmarks among all methods that accept RGB images. Moreover, it surpasses existing RGB-based 3D object detection methods on the SUN RGB-D dataset. On ScanNet, ImVoxelNet sets a new benchmark for multi-view 3D object detection.

While unlabelled data can be largely available and even abundant, the annotation process can be quite expensive and limiting. Under the assumption that some samples are more important for a given task than others, active learning targets the problem of identifying the most informative samples that one should acquire annotations for. In this work we propose a simple sample selection criterion that moves beyond the conventional reliance on model uncertainty as proxy to leverage new labels. By first accepting the model prediction and then judging its effect on the generalization error, we can better identify wrongly predicted samples. We also present a very efficient approximation to our criterion, providing a similarity-based interpretation. In addition to evaluating our method on the standard benchmarks of active learning, we consider the challenging yet realistic imbalanced data scenario. We show state-of-the-art results, especially on the imbalanced setting, and achieve better rates at identifying wrongly predicted samples than existing active learning methods. Our method is simple, model agnostic and relies on the current model status without the need for re-training from scratch.

This work studies Hyperspectral image (HSI) super-resolution (SR). HSI SR is characterized by high-dimensional data and a limited amount of training exam-ples. This raises challenges for training deep neural net-works that are known to be data hungry. This work ad-dresses this issue with two contributions. First, we observethat HSI SR and RGB image SR are correlated and developa novel multi-tasking network to train them jointly so thatthe auxiliary task RGB image SR can provide additionalsupervision and regulate the network training. Second,we extend the network to a semi-supervised setting so thatit can learn from datasets containing only low-resolutionHSIs. With these contributions, our method is able to learnhyperspectral image super-resolution from heterogeneousdatasets and lifts the requirement for having a large amountof HD HSI training samples. Extensive experiments onthree standard datasets show that our method outperformsexisting methods significantly and underpin the relevance ofour contributions.

The convolution operation is a central building block of neural network architectures widely used in computer vision. The size of the convolution kernels determines both the expressiveness of convolutional neural networks (CNN), as well as the number of learnable parameters. Increasing the network capacity to capture rich pixel relationships requires increasing the number of learnable parameters, often leading to overfitting and/or lack of robustness. In this paper, we propose a powerful novel building block, the hyper-convolution, which implicitly represents the convolution kernel as a function of kernel coordinates. Hyper-convolutions enable decoupling the kernel size, and hence its receptive field, from the number of learnable parameters. In our experiments, focused on challenging biomedical image segmentation tasks, we demonstrate that replacing regular convolutions with hyper-convolutions leads to more efficient architectures that achieve improved accuracy. Our analysis also shows that learned hyper-convolutions are naturally regularized, which can offer better generalization performance. We believe that hyper-convolutions can be a powerful building block in future neural network architectures solving computer vision tasks. We provide all of our code here: https://github.com/tym002/Hyper-Convolution

We present HybVIO, a novel hybrid approach for combining filtering-based visual-inertial odometry (VIO) with optimization-based SLAM. The core of our method is highly robust, independent VIO with improved IMU bias modeling, outlier rejection, stationarity detection, and feature track selection, which is adjustable to run on embedded hardware. Long-term consistency is achieved with a loosely-coupled SLAM module. In academic benchmarks, our solution yields excellent performance in all categories, especially in the real-time use case, where we outperform the current state-of-the-art. We also demonstrate the feasibility of VIO for vehicular tracking on consumer-grade hardware using a custom dataset, and show good performance in comparison to current commercial VISLAM alternatives.

Deep learning has driven remarkable accuracy increases in many computer vision problems. One ongoing challenge is how to achieve the greatest accuracy in cases where training data is limited. A second ongoing challenge is that trained models oftentimes do not generalize well even to new data that is subjectively similar to the training set. We address these challenges in a novel way, with the first-ever (to our knowledge) exploration of encoding human judgement about salient regions of images into the training data. We compare the accuracy and generalization of a state-of-the-art deep learning algorithm for a difficult problem in biometric presentation attack detection when trained on (a) original images with typical data augmentations, and (b) the same original images transformed to encode human judgement about salient image regions. The latter approach results in models that achieve higher accuracy and better generalization, decreasing the error of the LivDet-Iris 2020 winner from 29.78% to 16.37%, and achieving impressive generalization in a leave-one-attack-type-out evaluation scenario. This work opens a new area of study for how to embed human intelligence into training strategies for deep learning to achieve high accuracy and generalization in cases of limited training data.

There are two competing standards for self-supervised learning in action recognition from 3D skeletons. Su et al., 2020 used an auto-encoder architecture and an image reconstruction objective function to achieve state-of-the-art performance on the NTU60 C-View benchmark. Rao et al., 2020 used Contrastive learning in the latent space to achieve state-of-the-art performance on the NTU60 C-Sub benchmark. Here, we reconcile these disparate approaches by developing a taxonomy of self-supervised learning for action recognition. We observe that leading approaches generally use one of two types of objective functions: those that seek to reconstruct the input from a latent representation ("Attractive" learning) versus those that also try to maximize the representations distinctiveness ("Contrastive" learning). Independently, leading approaches also differ in how they implement these objective functions: there are those that optimize representations in the decoder output space and those which optimize representations in the network's latent space (encoder output). We find that combining these approaches leads to larger gains in performance and tolerance to transformation than is achievable by any individual method, leading to state-of-the-art performance on three standard action recognition datasets. We include links to our code and data.

A plethora of methods have been proposed to explain how deep neural networks reach their decisions but comparatively, little effort has been made to ensure that the explanations produced by these methods are objectively relevant. While several desirable properties for trustworthy explanations have been formulated, objective measures have been harder to derive. Here, we propose two new measures to evaluate explanations borrowed from the field of algorithmic stability: mean generalizability MeGe and relative consistency ReCo. We conduct extensive experiments on different network architectures, common explainability methods, and several image datasets to demonstrate the benefits of the proposed measures. In comparison to ours, popular fidelity measures are not sufficient to guarantee trustworthy explanations. Finally, we found that 1-Lipschitz networks produce explanations with higher MeGe and ReCo than common neural networks while reaching similar accuracy. This suggests that 1-Lipschitz networks are a relevant direction towards predictors that are more explainable and trustworthy.

"Wireframe" is a line segment based representation designed to well capture large-scale visual properties of regular, structural shaped man-made scenes surrounding us. Unlike the wireframes, conventional edges or line segments focus on all visible edges and lines without particularly distinguishing which of them are more salient to man-made structural information. Existing wireframe detection models rely on supervising the annotated data but do not explicitly pay attention to understand how to compose the structural shapes of the scene. In addition, we often face that many foreground objects occluding the background scene interfere with proper inference of the full scene structure behind them. To resolve these problems, we first time in the field, propose new conditional data generation and training that help the model understand how to ignore occlusion indicated by holes, such as foreground object regions masked out on the image. In addition, we first time combine GAN in the model to let the model better predict underlying scene structure even beyond large holes. We also introduce pseudo labeling to further enlarge the model capacity to overcome small-scale labeled data. We show qualitatively and quantitatively that our approach significantly outperforms previous works unable to handle holes, as well as improves ordinary detection without holes given.

High dynamic range (HDR) imaging is a highly challenging task since a large amount of information is lost due to the limitations of camera sensors. For HDR imaging, some methods capture multiple low dynamic range (LDR) images with altering exposures to aggregate more information. However, these approaches introduce ghosting artifacts when significant inter-frame motions are present. Moreover, although multi-exposure images are given, we have little information in severely over-exposed areas. Most existing methods focus on motion compensation, i.e., alignment of multiple LDR shots to reduce the ghosting artifacts, but they still produce unsatisfying results. These methods also rather overlook the need to restore the saturated areas. In this paper, we generate well-aligned multi-exposure features by reformulating a motion alignment problem into a simple brightness adjustment problem. In addition, we propose a coarse-to-fine merging strategy with explicit saturation compensation. The saturated areas are reconstructed with similar well-exposed content using adaptive contextual attention. We demonstrate that our method outperforms the state-of-the-art methods regarding qualitative and quantitative evaluations.

We present a novel technique for self-supervised video representation learning by: (a) decoupling the learning objective into two contrastive subtasks respectively emphasizing spatial and temporal features, and (b) performing it hierarchically to encourage multi-scale understanding. Motivated by their effectiveness in supervised learning, we first introduce spatial-temporal feature learning decoupling and hierarchical learning to the context of unsupervised video learning. We show by experiments that augmentations can be manipulated as regularization to guide the network to learn desired semantics in contrastive learning, and we propose a way for the model to separately capture spatial and temporal features at multiple scales. We also introduce an approach to overcome the problem of divergent levels of instance invariance at different hierarchies by modeling the invariance as loss weights for objective re-weighting. Experiments on downstream action recognition benchmarks on UCF101 and HMDB51 show that our proposed Hierarchically Decoupled Spatial-Temporal Contrast (HDC) makes substantial improvements over directly learning spatial-temporal features as a whole and achieves competitive performance when compared with other state-of-the-art unsupervised methods. Code will be made available.

Proxy-based metric learning losses are superior to pair-based losses due to their fast convergence and low training complexity. However, existing proxy-based losses focus on learning class-discriminative features while overlooking the commonalities shared across classes which are potentially useful in describing and matching samples. Moreover, they ignore the implicit hierarchy of categories in real-world datasets, where similar subordinate classes can be grouped together. In this paper, we present a framework that leverages this implicit hierarchy by imposing a hierarchical structure on the proxies and can be used with any existing proxy-based loss. This allows our model to capture both class-discriminative features and class-shared characteristics without breaking the implicit data hierarchy. We evaluate our method on five established image retrieval datasets such as In-Shop and SOP. Results demonstrate that our hierarchical proxy-based loss framework improves the performance of existing proxy-based losses, especially on large datasets which exhibit strong hierarchical structure.

This paper focuses on task recognition and action segmentation in weakly-labeled instructional videos, where only the ordered sequence of video-level actions is available during training. We propose a two-stream framework, which exploits semantic and temporal hierarchies to recognize top-level tasks in instructional videos. Further, we present a novel top-down weakly-supervised action segmentation approach, where the predicted task is used to constrain the inference of fine-grained action sequences. Experimental results on the popular Breakfast and Cooking 2 datasets show that our two-stream hierarchical task modeling significantly outperforms existing methods in top-level task recognition for all datasets and metrics. Additionally, using our task recognition framework in the proposed top-down action segmentation approach consistently improves the state of the art, while also reducing segmentation inference time by 80-90 percent.

Semi-supervised learning approaches have emerged as an active area of research to combat the challenge of obtaining large amounts of annotated data. Towards the goal of improving the performance of semi-supervised learning methods, we propose a novel framework, HIERMATCH, a semi-supervised approach that leverages hierarchical information to reduce labeling costs and performs as well as a vanilla semi-supervised learning method. Hierarchical information is often available as prior knowledge in the form of coarse labels (e.g., woodpeckers) for images with fine-grained labels (e.g., downy woodpeckers or golden-fronted woodpeckers). However, the use of supervision using coarse-category labels to improve semi-supervised techniques has not been explored. In the absence of fine-grained labels, HIERMATCH exploits the label hierarchy and uses coarse class labels as a weak supervisory signal. Additionally, HIERMATCH is a generic-approach to improve any semi-supervised learning framework, we demonstrate this using our results on recent state-of-the-art techniques MixMatch and FixMatch. We evaluate the efficacy of HIERMATCH on two benchmark datasets, namely CIFAR-100 and NABirds. HIERMATCH can reduce the usage of fine-grained labels by 50% on CIFAR-100 with only a marginal drop of 0.59% in top-1 accuracy as compared to MixMatch.

Pruning is an effective method to reduce the memory footprint and FLOPs associated with neural network models. However, existing structured pruning methods often result in significant accuracy degradation for moderate pruning levels. To address this problem, we introduce a new Hessian Aware Pruning (HAP) method coupled with a Neural Implant approach that uses second-order sensitivity as a metric for structured pruning. The basic idea is to prune insensitive components and to use a Neural Implant for moderately sensitive components, instead of completely pruning them. For the latter approach, the moderately sensitive components are replaced with a low-rank implant that is smaller and less computationally expensive than the original component. We use the relative Hessian trace to measure sensitivity, as opposed to the magnitude-based sensitivity metric commonly used in the literature. We test HAP for both computer vision tasks and natural language tasks, and we achieve new state-of-the-art results. Specifically,HAP achieves less than 0.1%/0.5% degradation on PreResNet29/ResNet50(CIFAR-10/ImageNet) with more than 70%/50% of parameters pruned. Meanwhile, HAP also achieves significantly better performance (up to 0.8% with 60% of parameters pruned) as compared to gradient-based method for head pruning on transformer-based models.

In this paper we introduce a novel method to estimate the head pose of people in single images starting from a small set of head keypoints. To this purpose, we propose a regression model that exploits keypoints computed automatically by 2D pose estimation algorithms and outputs the head pose represented by yaw, pitch, and roll. Our model is simple to implement and more efficient with respect to the state of the art -- faster in inference and smaller in terms of memory occupancy -- with comparable accuracy. Our method also provides a measure of the heteroscedastic uncertainties associated with the three angles, through an appropriately designed loss function; we show there is a correlation between error and uncertainty values, thus this extra source of information may be used in subsequent computational steps. As an example application, we address social interaction analysis in images: we propose an algorithm for a quantitative estimation of the level of interaction between people, starting from their head poses and reasoning on their mutual positions.

Superpixels serve as a powerful preprocessing tool in many computer vision tasks. By using superpixel representation, the number of image primitives can be largely reduced by orders of magnitudes. The majority of superpixel methods use handcrafted features, which usually do not translate well into strong adherence to object boundaries. A few recent superpixel methods have introduced deep learning into the superpixel segmentation process. However, none of these methods is able to produce superpixels in near real-time, which is crucial to the applicability of a superpixel method in practice. In this work, we propose a two-stage graph-based framework for superpixel segmentation. In the first stage, we introduce an efficient Deep Affinity Learning (DAL) network that learns pairwise pixel affinities by aggregating multi-scale information. In the second stage, we propose a highly efficient superpixel method called Hierarchical Entropy Rate Segmentation (HERS). Using the learned affinities from the first stage, HERS builds a hierarchical tree structure that can produce any number of highly adaptive superpixels instantaneously. We demonstrate, through visual and numerical experiments, the effectiveness and efficiency of our method compared to various state-of-the-art superpixel methods.