WACV 2022 Papers

Different methods can be applied to satellite images to derive an altitude map from a set of images. In this article we evaluate a set of representative methods from different approaches. We consider true multi-view stereo methods as well as pair-wise ones, classic methods and deep learning based ones, methods already in use on satellite images and others that were originally devised for close range imaging and are adapted to satellite imagery. While deep learning (DL) methods have taken over multi-view stereo reconstruction in the last years, this tendency has not fully reached satellite stereo pipelines that still largely rely on pair-wise classic algorithms. For the comparison, we set-up a framework that allows to interface a DL-based stereo method taken from the computer vision literature with a satellite stereo pipeline. For multi-view stereo algorithms we build on a recently proposed framework originally devised to apply Colmap method to satellite images. Methods are compared on several datasets that include sets of images taken within a few days and sets of images taken months apart. Results show that DL methods have, in general, a good generalization power. In particular, the use of the GANet DL method as the matching step in a pair-wise stereo pipeline is promising as it already performs better than the classic counterpart, even without a specific training.

We address representation learning for large-scale instance-level image retrieval. Apart from backbone, training pipelines and loss functions, popular approaches have focused on different spatial pooling and attention mechanisms, which are at the core of learning a powerful global image representation. There are different forms of attention according to the interaction of elements of the feature tensor (local and global) and the dimensions where it is applied (spatial and channel). Unfortunately, each study addresses only one or two forms of attention and applies it to different problems like classification, detection or retrieval. We present global-local attention module (GLAM), which is attached at the end of a backbone network and incorporates all four forms of attention: local and global, spatial and channel. We obtain a new feature tensor and, by spatial pooling, we learn a powerful embedding for image retrieval. Focusing on global descriptors, we provide empirical evidence of the interaction of all forms of attention and improve the state of the art on standard benchmarks.

Efficient road traffic monitoring is playing a fundamental role in successfully resolving traffic congestion in cities.Unmanned Aerial Vehicles (UAVs) or drones equipped with cameras are an attractive proposition to provide flexible and infrastructure-free traffic monitoring. However, real-time traffic monitoring from UAV imagery poses several challenges, due to the large image sizes and presence of non relevant targets. In this paper, we propose the AirCam-RTM framework that combines road segmentation and vehicle detection to focus only on relevant vehicles, which as a result, improves the monitoring performance by approximately 2x and provides approximately 18% accuracy improvement. Furthermore,through a real experimental setup we qualitatively evaluate the performance of the proposed approach, and also demonstrate how it can be used for real-time traffic monitoring and management using UAVs.

Deep Learning of neural networks has progressively become more prominent in healthcare with models reaching, or even surpassing, expert accuracy levels. However, these success stories are tainted by concerning reports on the lack of model transparency and bias against some medical conditions or patients' sub-groups. Explainable methods are considered the gateway to alleviate many of these concerns. In this study we demonstrate that the generated explanations are volatile to changes in model training that are perpendicular to the classification task and model structure. This raises further questions about trust in deep learning models for healthcare. Mainly, whether the models capture underlying causal links in the data or just rely on spurious correlations that are made visible via explanation methods. We demonstrate that the output of explainability methods on deep neural networks can vary significantly by changes of hyper-parameters, such as the random seed or how the training set is shuffled. We introduce a measure of explanation consistency which we use to highlight the identified problems on the MIMIC-CXR dataset. We find explanations of identical models but with different training setups have a low consistency: approximately 33% on average. On the contrary, kernel methods are robust against any orthogonal changes, with explanation consistency at 94%. We conclude that current trends in model explanation are not sufficient to mitigate the risks of deploying models in real life healthcare applications.

Convolutional neural networks typically perform poorly when the test (target domain) and training (source domain) data have significantly different distributions. While this problem can be mitigated by using the target domain data to align the source and target domain feature representations, the target domain data may be unavailable due to privacy concerns. Consequently, there is a need for methods that generalize well despite restricted access to target domain data during training. In this work, we propose an adversarial semantic hallucination approach (ASH), which combines a class-conditioned hallucination module and a semantic segmentation module. Since the segmentation performance varies across different classes, we design a semantic-conditioned style hallucination module to generate affine transformation parameters from semantic information in the segmentation probability maps of the source domain image. Unlike previous adaptation approaches, which treat all classes equally, ASH considers the class-wise differences. The segmentation module and the hallucination module compete adversarially, with the hallucination module generating increasingly "difficult" stylized images to challenge the segmentation module. In response, the segmentation module improves as it is trained with generated samples at an appropriate class-wise difficulty level. Our results on the Cityscapes and Mapillary benchmark datasets show that our method is competitive with state of the art work. Code is made available at https://github.com/gabriel-tjio/ASH.

We experimentally study the robustness of deep camera-LiDAR fusion architectures for 2D object detection in autonomous driving. First, we find that the fusion model is usually both more accurate, and more robust against single-source attacks than single-sensor deep neural networks. Furthermore, we show that without adversarial training, early fusion is more robust than late fusion, whereas the two perform similarly after adversarial training. However, we note that single-channel adversarial training of deep fusion is often detrimental even to robustness. Moreover, we observe cross-channel externalities, where single-channel adversarial training reduces robustness to attacks on the other channel. Additionally, we observe that the choice of adversarial model in adversarial training is critical: using attacks restricted to cars' bounding boxes is more effective in adversarial training and exhibits less significant cross-channel externalities. Finally, we find that joint-channel adversarial training helps mitigate many of the issues above, but does not significantly boost adversarial robustness.

In this paper, we explore open-domain sketch-to-photo translation, which aims to synthesize a realistic photo from a freehand sketch with its class label, even if the sketches of that class are missing in the training data. It is challenging due to the lack of training supervision and the large geometric distortion between the freehand sketch and photo domains. To synthesize the absent freehand sketches from photos, we propose a framework that jointly learns sketch-to-photo and photo-to-sketch generation. However, the generator trained from fake sketches might lead to unsatisfying results when dealing with sketches of missing classes, due to the domain gap between synthesized sketches and real ones. To alleviate this issue, we further propose a simple yet effective open-domain sampling and optimization strategy to "fool" the generator into treating fake sketches as real ones. Our method takes advantage of the learned sketch-to-photo and photo-to-sketch mapping of in-domain data and generalizes it to the open-domain classes. We validate our method on the Scribble and SketchyCOCO datasets. Compared with the recent competing methods, our approach shows impressive results in synthesizing realistic color, texture, and maintaining the geometric composition for various categories of open-domain sketches.

Unsupervised Domain Adaptation (UDA) is a key issue in visual recognition, as it allows to bridge different visual domains enabling robust performances in the real world. To date, all proposed approaches rely on human expertise to manually adapt a given UDA method (e.g. DANN) to a specific backbone architecture (e.g. ResNet). This dependency on handcrafted designs limits the applicability of a given approach in time, as old methods need to be constantly adapted to novel backbones. Existing Neural Architecture Search (NAS) approaches cannot be directly applied to mitigate this issue, as they rely on labels that are not available in the UDA setting. Furthermore, most NAS methods search for full architectures, which precludes the use of pre-trained models, essential in a vast range of UDA settings for reaching SOTA results. To the best of our knowledge, no prior work has addressed these aspects in the context of NAS for UDA. Here we tackle both aspects with an Adversarial Branch Architecture Search for UDA (ABAS): i. we address the lack of target labels by a novel data-driven ensemble approach for model selection; and ii. we search for an auxiliary adversarial branch, attached to a pre-trained backbone, which drives the domain alignment. We extensively validate ABAS to improve two modern UDA techniques, DANN and ALDA, on three standard visual recognition datasets (Office31, Office-Home and PACS). In all cases, ABAS robustly finds the adversarial branch architectures and parameters which yield best performances. https://github.com/lr94/abas

A recurring focus of the deep learning community is towards reducing the labeling effort. Data gathering and annotation using a search engine is a simple alternative to generating a fully human-annotated and human-gathered dataset. Although web crawling is very time efficient, some of the retrieved images are unavoidably noisy, i.e. incorrectly labeled. Designing robust algorithms for training on noisy data gathered from the web is an important research perspective that would render the building of datasets easier. In this paper we conduct a study to understand the type of label noise to expect when building a dataset using a search engine. We review the current limitations of state-of-the-art methods for dealing with noisy labels for image classification tasks in the case of web noise distribution. We propose a simple solution to bridge the gap with a fully clean dataset using Dynamic Softening of Out-of-distribution Samples (DSOS), which we design on corrupted versions of the CIFAR-100 dataset, and compare against state-of-the-art algorithms on the web noise perturbated MiniImageNet and Stanford datasets and on real label noise datasets: WebVision 1.0 and Clothing1M. Our work is fully reproducible https://git.io/JKGcj.

Remote sensing data is crucial for applications ranging from monitoring forest fires and deforestation to tracking urbanization. Most of these tasks require dense pixel-level annotations for the model to parse visual information from limited labeled data available for these satellite images. Due to the dearth of high-quality labeled training data in this domain, there is a need to focus on semi-supervised techniques. These techniques generate pseudo-labels from a small set of labeled examples which are used to augment the labeled training set. This makes it necessary to have a highly representative and diverse labeled training set. Therefore, we propose to use an active learning-based sampling strategy to select a highly representative set of labeled training data. We demonstrate our proposed method's effectiveness on two existing semantic segmentation datasets containing satellite images: UC Merced Land Use Classification Dataset and DeepGlobe Land Cover Classification Dataset. We report a 27% improvement in mIoU with as little as 2% labeled data using active learning sampling strategies over randomly sampling the small set of labeled training data.

To get something done, humans perform a sequence of actions dictated by a goal. So, predicting the next action in the sequence becomes easier once we know the goal that guides the entire activity. We present an action anticipation model that uses goal information in an effective manner. Specifically, we use a latent goal representation as a proxy for the "real goal" of the sequence and use this goal information when predicting the next action. We design a model to compute the latent goal representation from the observed video and use it to predict the next action. We also exploit two properties of goals to propose new losses for training the model. First, the effect of the next action should be closer to the latent goal than the observed action, termed as "goal closeness". Second, the latent goal should remain consistent before and after the execution of the next action which we coined as "goal consistency". Using this technique, we obtain state-of-the-art action anticipation performance on scripted datasets 50Salads and Breakfast that have predefined goals in all their videos. We also evaluate the latent goal-based model on EPIC-KITCHENS55 which is an unscripted dataset with multiple goals being pursued simultaneously. Even though this is not an ideal setup for using latent goals, our model is able to predict the next noun better than existing approaches on both seen and unseen kitchens in the test set.

Recent advances in transformer-based models have drawn attention to exploring these techniques in medical image segmentation, especially in conjunction with the U-Net model (or its variants), which has shown great success in medical image segmentation, under both 2D and 3D settings. Current 2D based methods either directly replace convolutional layers with pure transformers or consider a transformer as an additional intermediate encoder between the encoder and decoder of U-Net. However, these approaches only consider the attention encoding within one single slice and do not utilize the axial-axis information naturally provided by a 3D volume. In the 3D setting, convolution on volumetric data and transformers both consume large GPU memory. One has to either downsample the image or use cropped local patches to reduce GPU memory usage, which limits its performance. In this paper, we propose Axial Fusion Transformer UNet (AFTer-UNet), which takes both advantages of convolutional layers' capability of extracting detailed features and transformers' strength on long sequence modeling. It considers both intra-slice and inter-slice long-range cues to guide the segmentation. Meanwhile, it has fewer parameters and takes less GPU memory to train than the previous transformer-based models. Extensive experiments on three multi-organ segmentation datasets demonstrate that our method outperforms current state-of-the-art methods.

StyleGANs have shown impressive results on data generation and manipulation in recent years, thanks to its disentangled style latent space. A lot of efforts have been made in inverting a pre-trained generator, where an encoder is trained ad hoc after the generator is trained in a two-stage fashion. In this paper, we focus on style-based generators asking a scientific question: Does forcing such a generator to reconstruct real data lead to more disentangled latent space and make the inversion process from image to latent space easy? We describe a new methodology to train a style-based autoencoder where the encoder and generator are optimized end-to-end. We show that our proposed model consistently outperforms baselines in terms of image inversion and generation quality.

Deep Neural Networks (DNNs) have been shown to be vulnerable to adversarial examples, which are slightly perturbed input images which lead DNNs to make wrong predictions. To protect from such examples, various defense strategies have been proposed. A very recent defense strategy for detecting adversarial examples, that has been shown to be robust to current attacks, is to check for intrinsic context consistencies in the input data, where context refers to various relationships (e.g., object-to-object co-occurrence relationships) in images. In this paper, we show that even context consistency checks can be brittle to properly crafted adversarial examples and to the best of our knowledge, we are the first to do so. Specifically, we propose an adaptive framework to generate examples that subvert such defenses, namely, Adversarial attacks against object Detection that evade Context consistency checks (ADC). In ADC, we formulate a joint optimization problem which has two attack goals, viz., (i) fooling the object detector and (ii) evading the context consistency check system, at the same time. Experiments on both PASCAL VOC and MS COCO datasets show that examples generated with ADC fool the object detector with a success rate of over 85% in most cases, and at the same time evade the recently proposed context consistency checks, with a bypassing rate of over 80% in most cases. Our results suggest that how to robustly model context and check its consistency, is still an open problem.

Estimating camera pose from a single image is a fundamental problem in computer vision. Existing methods for solving this task fall into two distinct categories, which we refer to as direct and indirect. Direct methods, such as PoseNet, regress pose from the image as a fixed function, for example using a feed-forward convolutional network. Such methods are desirable because they are deterministic and run in constant time. Indirect methods for pose regression are often non-deterministic, with various external dependencies such as image retrieval and hypothesis sampling. We propose a direct method that takes inspiration from structure-based approaches to incorporate explicit 3D constraints into the network. Our approach maintains the desirable qualities of other direct methods while achieving much lower error in general. Code is available https://github.com/mvrl/structure-aware-pose-estimation.

Anomaly detection is a one-class classification (OCC) problem where the methods learn either a generative model of the inlier class (e.g., in the variants of kernel principal component analysis) or a decision boundary to encapsulate the inlier class (e.g., in the one-class variants of the support vector machine). Learning schemes for OCC typically rely on training data solely from the inlier class, but some recent approaches have proposed semi-supervised extensions, e.g., variants of semi-supervised anomaly detection that also leverage a small amount of training data from outlier classes. Other recent methods extend existing principles to employ deep neural network (DNN) modeling that relies on learning (for the inlier class) either latent-space distributions or autoencoders, but not both. We propose a novel semi-supervised variational formulation, leveraging generalized-Gaussian models leading to data-adaptive, robust, and uncertainty-aware distribution modeling in both latent space and image space. For variational learning, we propose a novel reparameterization for sampling from the latent-space generalized-Gaussian to enable backpropagation-based optimization. Results on several public image sets show the benefits of our method over state of the art.

We propose a novel framework for comparing 3D human shapes under the change of shape and pose. This problem is challenging since 3D human shapes vary significantly across subjects and body postures. We solve this problem by using a Riemannian approach. Our core contribution is the mapping of the human body surface to the space of metrics and normals. We equip this space with a family of Riemannian metrics, called Ebin (or DeWitt) metrics. We treat a human body surface as a point in a "shape space" equipped with a family of Riemmanian metrics. The family of metrics is invariant under rigid motions and reparametrizations; hence it induces a metric on the "shape space" of surfaces. Using the alignment of human bodies with a given template, we show that this family of metrics allows us to distinguish the changes in shape and pose. The proposed framework has several advantages. First, we define a family of metrics with desired invariant properties for the comparison of human shape. Second, we present an efficient framework to compute geodesic paths between human shape given the chosen metric. Third, this framework provides some basic tools for statistical shape analysis of human body surfaces. Finally, we demonstrate the utility of the proposed framework in pose and shape retrieval of human body.

Few-shot semantic segmentation addresses the learning task in which only few images with ground truth pixel-level labels are available for the novel classes of interest. One is typically required to collect a large mount of data (i.e., base classes) with such ground truth information, followed by meta-learning strategies to address the above learning task. When only image-level semantic labels can be observed during both training and testing, it is considered as an even more challenging task of weakly supervised few-shot semantic segmentation. To address this problem, we propose a novel meta-learning framework, which predicts pseudo pixel-level segmentation masks from a limited amount of data and their semantic labels. More importantly, our learning scheme further exploits the produced pixel-level information for query image inputs with segmentation guarantees. Thus, our proposed learning model can be viewed as a pixel-level meta-learner. Through extensive experiments on benchmark datasets, we show that our model achieves satisfactory performances under fully supervised settings, yet performs favorably against state-of-the-art methods under weakly supervised settings.

Anomaly detection in videos has been attracting an increasing amount of attention. Despite the competitive performance of recent methods on benchmark datasets, they typically lack desirable features such as modularity, cross-domain adaptivity, interpretability, and real-time anomalous event detection. Furthermore, current state-of-the-art approaches are evaluated using the standard instance-based detection metric by considering video frames as independent instances, which is not ideal for video anomaly detection. Motivated by these research gaps, we propose a modular and unified approach to the online video anomaly detection and localization problem, called MOVAD, which consists of a novel transfer learning based plug-and-play architecture, a sequential anomaly detector, a mathematical framework for selecting the detection threshold, and a suitable performance metric for real-time anomalous event detection in videos. Extensive performance evaluations on benchmark datasets show that the proposed framework significantly outperforms the current state-of-the-art approaches.