WACV 2022 Papers

Uncertainty estimation in medical image registration enables surgeons to evaluate the operative risk based on the trustworthiness of the registered image data thus of paramount importance for practical clinical applications. Despite the recent promising results obtained with deep unsupervised learning-based registration methods, reasoning about uncertainty of unsupervised registration models remains largely unexplored. In this work, we propose a predictive module to learn the registration and uncertainty in correspondence simultaneously. Our framework introduces empirical randomness and registration error based uncertainty prediction. We systematically assess the performances on two MRI datasets with different ensemble paradigms. Experimental results highlight that our proposed framework significantly improves the registration accuracy and uncertainty compared with the baseline.

Fully Convolutional Neural Networks (FCNNs) with contracting and expanding paths have shown prominence for the majority of medical image segmentation applications since the past decade. In FCNNs, the encoder plays an integral role by learning both global and local features and contextual representations which can be utilized for semantic output prediction by the decoder. Despite their success, the locality of convolutional layers in FCNNs, limits the capability of learning long-range spatial dependencies. Inspired by the recent success of transformers for Natural Language Processing (NLP) in long-range sequence learning, we reformulate the task of volumetric (3D) medical image segmentation as a sequence-to-sequence prediction problem. We introduce a novel architecture, dubbed as UNEt TRansformers (UNETR), that utilizes a transformer as the encoder to learn sequence representations of the input volume and effectively capture the global multi-scale information, while also following the successful "U-shaped" network design for the encoder and decoder. The transformer encoder is directly connected to a decoder via skip connections at different resolutions to compute the final semantic segmentation output. We have validated the performance of our method on the Multi Atlas Labeling Beyond The Cranial Vault (BTCV) dataset for multi-organ segmentation and the Medical Segmentation Decathlon (MSD) dataset for brain tumor and spleen segmentation tasks. Our benchmarks demonstrate new state-of-the-art performance on the BTCV leaderboard.

Text entry for mobile devices nowadays is an equally crucial and time-consuming task, with no practical solution available for natural typing speeds without extra hardware. In this paper, we introduce a real-time method that is a significant step towards enabling touch typing on arbitrary flat surfaces (e.g., tables). The method employs only a simple video camera, placed in front of the user on the flat surface --- at an angle practical for mobile usage. To achieve this, we adopt a classification framework, based on the observation that, in touch typing, similar hand configurations imply the same typed character across users. Importantly, this approach allows the convenience of un-calibrated typing, where the hand positions, with respect to the camera and each other, are not dictated. To improve accuracy, we propose a Language Processing scheme, which corrects the typed text and is specifically designed for real-time performance and integration with the vision-based signal. To enable feasible data collection and training, we propose a self-refinement approach that allows training on unlabeled flat-surface-typing footage; A network trained on (labeled) keyboard footage labels flat-surface videos using dynamic time warping, and is trained on them, in an Expectation Maximization (EM) manner. Using these techniques, we introduce the TypingHands26 Dataset, comprising videos of 26 different users typing on a keyboard, and 10 users typing on a flat surface, labeled at the frame level. We validate our approach and present a single camera-based system with character-level accuracy of 93.5% on average for known users, and 85.7% for unknown ones, outperforming pose-estimation-based methods by a large margin, despite performing at natural typing speeds of up to 80 Words Per Minute. Our method is the first to rely on a simple camera alone, and runs in interactive speeds, while still maintaining accuracy comparable to systems employing non-commodity equipment.

We present a novel approach for oriented object detection, named TricubeNet, which localizes oriented objects using visual cues (i.e., heatmap) instead of oriented box offsets regression. We represent each object as a 2D Tricube kernel and extract bounding boxes using simple image-processing algorithms. Our approach is able to (1) obtain well-arranged boxes from visual cues, (2) solve the angle discontinuity problem, and (3) can save computational complexity due to our anchor-free modeling. To further boost the performance, we propose some effective techniques for size-invariant loss, reducing false detections, extracting rotation-invariant features, and heatmap refinement. To demonstrate the effectiveness of our TricubeNet, we experiment on various tasks for weakly-occluded oriented object detection: detection in an aerial image, densely packed object image, and text image. The extensive experimental results show that our TricubeNet is quite effective for oriented object detection. Code is available at https://github.com/qjadud1994/TricubeNet.

Deep visual models are known to be vulnerable to adversarial attacks. The last few years have seen numerous techniques to compute adversarial inputs for these models. However, there are still under-explored avenues in this critical research direction. Among those is the estimation of adversarial textures for 3D models in an end-to-end optimization scheme. In this paper, we propose such a scheme to generate adversarial textures for 3D models that are highly transferable and invariant to different camera views and lighting conditions. Our method makes use of neural rendering with explicit control over the model texture and background. We ensure transferability of the adversarial textures by employing an ensemble of robust and non-robust models. Our technique utilizes 3D models as a proxy to simulate closer to real-life conditions, in contrast to conventional use of 2D images for adversarial attacks. We show the efficacy of our method with extensive experiments.

Human pose information is a critical component in many downstream image processing tasks, such as activity recognition and motion tracking. Likewise, a pose estimator for the illustrated character domain would provide a valuable prior for assistive content creation tasks, such as reference pose retrieval and automatic character animation. But while modern data-driven techniques have substantially improved pose estimation performance on natural images, little work has been done for illustrations. In our work, we bridge this domain gap by efficiently transfer-learning from both domain-specific and task-specific source models. Additionally, we upgrade and expand an existing illustrated pose estimation dataset, and introduce two new datasets for classification and segmentation subtasks. We then apply the resultant state-of-the-art character pose estimator to solve the novel task of pose-guided illustration retrieval. All data, models, and code will be made publicly available.

Player detection lays the foundation for many applications in the field of sports analytics including player recognition, player tracking, and activity detection. In this work we study player detection in continuous long shot broadcast videos. Broadcast match videos are easy to obtain, and detection on these videos is much more challenging. We propose a transductive approach for player detection that treats it as a domain adaptation problem. We show that instance-level domain labels are significant for sufficient adaptation in the case of soccer broadcast videos. An efficient multi-model greedy labelling scheme based on visual features is proposed to annotate domain labels on bounding box predictions made by our inductive model. We use reliable instances from the inductive model inferences to train a transductive copy of the model. We create and release a fully annotated player detection dataset comprising soccer broadcast videos from the FIFA 2018 World Cup matches to evaluate our method. Our method shows significant improvements in player detection to the baseline and existing state-of-the-art methods on our dataset. We show, on average, a 16 point improvement in mAP for soccer broadcast videos by annotating domain labels for around a 100 samples per video.

The choice of a loss function is an important factor when training neural networks for image restoration problems, such as single image super resolution. The loss function should encourage natural and perceptually pleasing results. A popular choice for a loss is a pre-trained network, such as VGG, which is used as a feature extractor for computing the difference between restored and reference images. However, such an approach has multiple drawbacks: it is computationally expensive, requires regularization and hyper-parameter tuning, and involves a large network trained on an unrelated task. Furthermore, it has been observed that there is no single loss function that works best across all applications and across different datasets. In this work, we instead propose to train a set of loss functions that are application specific in nature. Our loss function comprises a series of discriminators that are trained to detect and penalize the presence of application-specific artefacts. We show that a single natural image and corresponding distortions are sufficient to train our feature extractor that outperforms state-of-the-art loss functions in applications like single image super resolution, denoising, and JPEG artefact removal. Finally, we conclude that an effective loss function does not have to be a good predictor of perceived image quality, but instead needs to be specialized in identifying the distortions for a given restoration method.

Zero-shot classification is the task of learning predictors for classes not seen during training. A practical way to deal with the lack of annotations for the target categories is to encode not only the inputs (images) but also the outputs (object classes) into a suitable representation space. We can use these representations to measure the degree at which images and categories agree by fitting a compatibility measure using the information available during training. One way to define such a measure is by a two step process in which we first project the elements of either space visual or semantic) onto the other and then compute a similarity score in the target space. Although projections onto the visual space has shown better general performance, little attention has been paid to the degree at which the visual and semantic information contribute to the final predictions. In this paper, we build on this observation and propose two different formulations that allow us to explicitly trade-off the relative importance of the visual and semantic spaces for classification in a zero-shot setting. Our formulations are based on redefinition of the similarity scoring and loss function used to learn the projections. Experiments on six different datasets show that our approach lead to improve performance compared to similar methods. Moreover, combined with synthetic features, our approach competes favorably with the state of the art on both the standard and generalized settings.

We all have experienced the difficulty of designing appropriate neural architectures due to the lack of general principles and best practices. The game changer might be touse Neural Architecture Search (NAS) where a machine does all the hard work for us based on the data at its disposal. Invarious problems and in particular in classification, architectures designed by NAS outperform or compete with the best manual network designs in terms of accuracy, size, memory footprint and FLOPs. That said, previous studies focus ondeveloping NAS algorithms for clean high quality data, a restrictive and somewhat unrealistic assumption. In this paper, focusing on the differentiable NAS algorithms, we show that vanilla NAS algorithms suffer from a performance loss if class labels are noisy. To combat this issue, we propose tomake use of the principle of information bottleneck as a regularizer. This leads us to develop a noise injecting operation that is included during the learning process, preventing the network from learning from noisy samples. Our empirical evaluations show that the noise injecting operation does not degrade the performance of the NAS algorithm if the data is indeed clean. In contrast, if the data is noisy, the architecture learned by our algorithm comfortably outperforms algorithms specifically equipped with sophisticated mechanisms to learn in the presence of label noise. In contrast to many algorithms designed to work in the presence of noisylabels, prior knowledge about the properties of the noise and its characteristics are not required for our algorithm.

This paper addresses bioprosthetic heart valve (BHV) durability estimation via computer vision (CV)-based analyses of the visual symmetry of valve leaflet motion. BHVs are routinely implanted in patients suffering from valvular heart diseases. Valve designs are rigorously tested using cardiovascular equipment, but once implanted, more than 50% of BHVs encounter a structural failure within 15 years. We investigate the correlation between the visual dynamic symmetry of BHV leaflets and the functional symmetry of the valves. We hypothesize that an asymmetry in the valve leaflet motion will generate an asymmetry in the flow patterns, resulting in added local stress and forces on some of the leaflets, which can accelerate the failure of the valve. We propose two different pair-wise leaflet symmetry scores based on the diagonals of orthogonal projection matrices (DOPM) and on dynamic time warping (DTW), computed from videos recorded during pulsatile flow tests. We compare the symmetry score profiles with those of fluid dynamic parameters (velocity and vorticity values) at the leaflet borders, obtained from valve-specific numerical simulations. Experiments on four cases that include three different tricuspid BHVs yielded promising results, with the DTW scores showing a good coherence with respect to the simulations. With a link between visual and functional symmetries established, this approach paves the way towards BHV durability estimation using CV techniques.

Poisoning attacks on machine learning systems compromise the model performance by deliberately injecting malicious samples in the training dataset to influence the training process. Prior works focus on either availability attacks (i.e., lowering the overall model accuracy) or integrity attacks (i.e., enabling specific instance-based backdoor). In this paper, we advance the adversarial objectives of the availability attacks to a per-class basis, which we refer to as class-oriented poisoning attacks. We demonstrate that the proposed attack is capable of forcing the corrupted model to predict in two specific ways: (i) classify unseen new images to a targeted "supplanter" class, and (ii) misclassify images from a "victim" class while maintaining the classification accuracy on other non-victim classes. To maximize the adversarial effect as well as reduce the computational complexity of poisoned data generation, we propose a gradient-based framework that crafts poisoning images with carefully manipulated feature information for each scenario. Using newly defined metrics at the class level, we demonstrate the effectiveness of the proposed class-oriented poisoning attacks on various models (e.g., LeNet-5, Vgg-9, and ResNet-50) over a wide range of datasets (e.g., MNIST, CIFAR-10, and ImageNet-ILSVRC2012) in an end-to-end training setting.

Visual event perception tasks such as action localization have primarily focused on supervised learning settings under a static observer, i.e., the camera is static and cannot be controlled by an algorithm. They are often restricted by the quality, quantity, and diversity of annotated training data and do not often generalize to out-of-domain samples. In this work, we tackle the problem of active action localization where the goal is to localize an action while controlling the geometric and physical parameters of an active camera to keep the action in the field of view without training data. We formulate an energy-based mechanism that combines predictive learning and reactive control to perform active action localization without rewards, which can be sparse or non-existent in real-world environments. We perform extensive experiments in both simulated and real-world environments on two tasks - active object tracking and active action localization. We demonstrate that the proposed approach can generalize to different tasks and environments in a streaming fashion, requiring only a single pass through the video, working in real-time. We show that the proposed approach outperforms unsupervised baselines and obtains competitive performance compared to those trained with reinforcement learning.

Advancements in adaptive object detection can lead to tremendous improvements in applications like autonomous navigation, as they alleviate the distributional shifts along the detection pipeline. Prior works adopt adversarial learning to align image features at global and local levels, yet the instance-specific misalignment persists. Also, adaptive object detection remains challenging due to visual diversity in background scenes and intricate combinations of objects. Motivated by structural importance, we aim to attend prominent instance-specific regions, overcoming the feature misalignment issue. We propose a novel resIduaL seLf-attentive featUre alignMEnt ( ILLUME ) method for adaptive object detection. ILLUME comprises Self-Attention Feature Map (SAFM) module that enhances structural attention to object-related regions and thereby generates domain invariant features. Our approach significantly reduces the domain distance with the improved feature alignment of the instances. Qualitative results demonstrate the ability of ILLUME to attend important object instances required for alignment. Experimental results on several benchmark datasets show that our method outperforms the existing state-of-the-art approaches.

We propose a novel solution for the task of video panoptic segmentation, that simultaneously predicts pixel-level semantic and instance segmentation and generates clip-level instance tracks. Our network, named VPS-Transformer, with a hybrid architecture based on the state-of-the-art panoptic segmentation network Panoptic-DeepLab, combines a convolutional architecture for single-frame panoptic segmentation and a novel video module based on an instantiation of the pure Transformer block. The Transformer, equipped with attention mechanisms, models spatio-temporal relations between backbone output features of current and past frames for more accurate and consistent panoptic estimates. As the pure Transformer block introduces large computation overhead when processing high resolution images, we propose a few design changes for a more efficient compute. We study how to aggregate information more effectively over the space-time volume and we compare several variants of the Transformer block with different attention schemes. Extensive experiments on the Cityscapes-VPS dataset demonstrate that our best model improves the temporal consistency and video panoptic quality by a margin of 2.2%, with little extra computation.

Over recent years, a myriad of novel convolutional network architectures have been developed to advance state-of-the-art performance on challenging recognition tasks. As computational resources improve, a great deal of effort has been placed on efficiently scaling up existing designs and generating new architectures with Neural Architecture Search (NAS) algorithms. While network topology has proven to be a critical factor for model performance, we show that significant gains are being left on the table by keeping topology static at inference-time. Due to challenges such as scale variation, we should not expect static models configured to perform well across a training dataset to be optimally configured to handle all test data. In this work, we expose the exciting potential of inference-time-dynamic models. We show that by allowing just four layers to dynamically change configuration at inference-time, off-the-shelf models like ResNet-50 have an upper bound accuracy of over 95% on ImageNet. This level of performance currently exceeds that of models with over 20x more parameters and significantly more complex training procedures. While this upper bound of performance may be practically difficult to achieve for a real dynamic model, it indicates a significant source of untapped potential for current models.

In many computer vision classification tasks, class priors at test time often differ from priors on the training set. In the case of such prior shift, classifiers must be adapted correspondingly to maintain close to optimal performance. This paper analyzes methods for adaptation of probabilistic classifiers to new priors and for estimating new priors on an unlabeled test set. We propose a novel method to address a known issue of prior estimation methods based on confusion matrices, where inconsistent estimates of decision probabilities and confusion matrices lead to negative values in the estimated priors. Experiments on fine-grained image classification datasets provide insight into the best practice of prior shift estimation and classifier adaptation, and show that the proposed method achieves state-of-the-art results in prior adaptation. Applying the best practice to two tasks with naturally imbalanced priors, learning from web-crawled images and plant species classification, increased the recognition accuracy by 1.1% and 3.4% respectively.

In this work we present a method that combines tensor-based face modelling and analysis and non-rigid structure-from-motion (NRSFM). The core idea is to see that the conventional tensor formulation for the face structure and expression analysis can be utilised while the structure component can be directly analysed as the non-rigid structure-from-motion problem. To the NRSFM problem part we further present a novel prior-free approach that factorises the 2D input shapes into affine projection matrices, rank-one 3D affine basis shapes, and the basis shape coefficients. The linear combination of the basis shapes thus yields the recovered 3D shapes upto an affine transformation. In contrast to most works in literature, no calibration information of the cameras or structure prior is required. Experiments on challenging face datasets show that our method, with and without the metric upgrade, is accurate and fast when compared to the state-of-the-art and is well suitable for dense reconstruction and face editing.

Few-shot learning addresses the challenge of learning how to address novel tasks given not just limited supervision but limited data as well. An attractive solution is synthetic data generation. However, most such methods are overly sophisticated, focusing on high-quality, realistic data in the input space. It is unclear whether adapting them to the few-shot regime and using them for the downstream task of classification is the right approach. Previous works on synthetic data generation for few-shot classification focus on exploiting complex models, e.g. a Wasserstein GAN with multiple regularizers or a network that transfers latent diversities from known to novel classes. We follow a different approach and investigate how a simple and straightforward synthetic data generation method can be used effectively. We make two contributions, namely we show that: (1) using a simple loss function is more than enough for training a feature generator in the few-shot setting; and (2) learning to generate tensor features instead of vector features is superior. Extensive experiments on miniImagenet, CUB and CIFAR-FS datasets show that our method sets a new state of the art, outperforming more sophisticated few-shot data augmentation methods. The source code can be found at https://github.com/MichalisLazarou/TFH_fewshot.