CVPR 2021 Papers

Deep neural networks have amply demonstrated their prowess but estimating the reliability of their predictions remains challenging. Deep Ensembles are widely considered as being one of the best methods for generating uncertainty estimates but are very expensive to train and evaluate. MC-Dropout is another popular alternative, which is less expensive, but also less reliable. Our central intuition is that there is a continuous spectrum of ensemble-like models of which MC-Dropout and Deep Ensembles are extreme examples. The first uses effectively infinite number of highly correlated models while the second relies on a finite number of independent models. To combine the benefits of both, we introduce Masksembles. Instead of randomly dropping parts of the network as in MC-dropout, Masksemble relies on a fixed number of binary masks, which are parameterized in a way that allows to change correlations between individual models. Namely, by controlling the overlap between the masks and their density one can choose the optimal configuration for the task at hand. This leads to a simple and easy to implement method with performance on par with Ensembles at a fraction of the cost. We experimentally validate Masksembles on two widely used datasets, CIFAR10 and ImageNet.

We introduce Mask-ToF, a method to reduce flying pixels (FP) in time-of-flight (ToF) depth captures. FPs are pervasive artifacts which occur around depth edges, where light paths from both an object and its background are integrated over the aperture. This light mixes at a sensor pixel to produce erroneous depth estimates, which can adversely affect downstream 3D vision tasks. Mask-ToF starts at the source of these FPs, learning a microlens-level occlusion mask which effectively creates a custom-shaped sub-aperture for each sensor pixel. This modulates the selection of foreground and background light mixtures on a per-pixel basis and thereby encodes scene geometric information directly into the ToF measurements. We develop a differentiable ToF simulator to jointly train a convolutional neural network to decode this information and produce high-fidelity, low-FP depth reconstructions. We test the effectiveness of Mask-ToF on a simulated light field dataset and validate the method with an experimental prototype. To this end, we manufacture the learned amplitude mask and design an optical relay system to virtually place it on a high-resolution ToF sensor. We find that Mask-ToF generalizes well to real data without retraining, cutting FP counts in half.

Semi-supervised generative learning (SSGL) makes use of unlabeled data to achieve a trade-off between the data collection/annotation effort and generation performance, when adequate labeled data are not available. Learning precise class semantics is crucial for class-conditional image synthesis with limited supervision. Toward this end, we propose a semi-supervised Generative Adversarial Network with a Mask-Embedded Discriminator, which is referred to as MED-GAN. By incorporating a mask embedding module, the discriminator features are associated with spatial information, such that the focus of the discriminator can be limited in the specified regions when distinguishing between real and synthesized images. A generator is enforced to synthesize the instances holding more precise class semantics in order to deceive the enhanced discriminator. Also benefiting from mask embedding, region-based semantic regularization is imposed on the discriminator feature space, and the degree of separation between real and fake classes and among object categories can thus be increased. This eventually improves class-conditional distribution matching between real and synthesized data. In the experiments, the superior performance of MED-GAN demonstrates the effectiveness of mask embedding and associated regularizers in facilitating SSGL.

We propose Mask Guided (MG) Matting, a robust matting framework that takes a general coarse mask as guidance. MG Matting leverages a network (PRN) design which encourages the matting model to provide self-guidance to progressively refine the uncertain regions through the decoding process. A series of guidance mask perturbation operations are also introduced in the training to further enhance its robustness to external guidance. We show that PRN can generalize to unseen types of guidance masks such as trimap and low-quality alpha matte, making it suitable for various application pipelines. In addition, we revisit the foreground color prediction problem for matting and propose a surprisingly simple improvement to address the dataset issue. Evaluation on real and synthetic benchmarks shows that MG Matting achieves state-of-the-art performance using various types of guidance inputs. Code and models are available at https://github.com/yucornetto/MGMatting.

The domain of Embodied AI has recently witnessed substantial progress, particularly in navigating agents within their environments. These early successes have laid the building blocks for the community to tackle tasks that require agents to actively interact with objects in their environment. Object manipulation is an established research domain within the robotics community and poses several challenges including manipulator motion, grasping and long-horizon planning, particularly when dealing with oft-overlooked practical setups involving visually rich and complex scenes, manipulation using mobile agents (as opposed to tabletop manipulation), and generalization to unseen environments and objects. We propose a framework for object manipulation built upon the physics-enabled, visually rich AI2-THOR framework and present a new challenge to the Embodied AI community known as ArmPointNav. This task extends the popular point navigation task to object manipulation and offers new challenges including 3D obstacle avoidance, manipulating objects in the presence of occlusion, and multi-object manipulation that necessitates long term planning. Popular learning paradigms that are successful on PointNav challenges show promise, but leave a large room for improvement.

Neural network pruning is an essential approach for reducing the computational complexity of deep models so that they can be well deployed on resource-limited devices. Compared with conventional methods, the recently developed dynamic pruning methods determine redundant filters variant to each input instance which achieves higher acceleration. Most of the existing methods discover effective sub-networks for each instance independently and do not utilize the relationship between different inputs. To maximally excavate redundancy in the given network architecture, this paper proposes a new paradigm that dynamically removes redundant filters by embedding the manifold information of all instances into the space of pruned networks (dubbed as ManiDP). We first investigate the recognition complexity and feature similarity between images in the training set. Then, the manifold relationship between instances and the pruned sub-networks will be aligned in the training procedure. The effectiveness of the proposed method is verified on several benchmarks, which shows better performance in terms of both accuracy and computational cost compared to the state-of-the-art methods. For example, our method can reduce 55.3% FLOPs of ResNet-34 with only 0.57% top-1 accuracy degradation on ImageNet.

We present Magic Layouts; a method for parsing screenshots or hand-drawn sketches of user interface (UI) layouts. Our core contribution is to extend existing detectors to exploit a learned structural prior for UI designs, enabling robust detection of UI components; buttons, text boxes and similar. Specifically we learn a prior over mobile UI layouts, encoding common spatial co-occurrence relationships between different UI components. Conditioning region proposals using this prior leads to performance gains on UI layout parsing for both hand-drawn UIs and app screenshots, which we demonstrate within the context an interactive application for rapidly acquiring digital prototypes of user experience (UX) designs.

The performance of face recognition system degrades when the variability of the acquired faces increases. Prior work alleviates this issue by either monitoring the face quality in pre-processing or predicting the data uncertainty along with the face feature. This paper proposes MagFace, a category of losses that learn a universal feature embedding whose magnitude before normalization can measure with the quality of the given face. Under the new loss, it can be proven that the magnitude of the feature embedding monotonically increases if the subject is more likely to be recognized. In addition, MagFace introduces an adaptive mechanism to learn a well-structured within-class feature distributions by pushing easy samples to class centers while pushing hard samples away. This prevents models from overfitting on noisy low-quality samples and improves face recognition in the wild. Extensive experiments conducted on face recognition, quality assessments as well as clustering have demonstrated the effectiveness of MagFace over state-of-the-arts. The code is available at https://github.com/IrvingMeng/MagFace.

Deepfakes raised serious concerns on the authenticity of visual contents. Prior works revealed the possibility to disrupt deepfakes by adding adversarial perturbations to the source data, but we argue that the threat has not been eliminated yet. This paper presents MagDR, a mask-guided detection and reconstruction pipeline for defending deepfakes from adversarial attacks. MagDR starts with a detection module that defines a few criteria to judge the abnormality of the output of deepfakes, and then uses it to guide an learnable reconstruction procedure. Adaptive masks are extracted to capture the change in local facial regions. In experiments, MagDR defends three main tasks of deepfakes, and the learned reconstruction pipeline transfers across input data, showing promising performance in defending both black-box and white-box attacks.

We present MaX-DeepLab, the first end-to-end model for panoptic segmentation. Our approach simplifies the current pipeline that depends heavily on surrogate sub-tasks and hand-designed components, such as box detection, non-maximum suppression, thing-stuff merging, etc. Although these sub-tasks are tackled by area experts, they fail to comprehensively solve the target task. By contrast, our MaX-DeepLab directly predicts class-labeled masks with a mask transformer, and is trained with a panoptic quality inspired loss via bipartite matching. Our mask transformer employs a dual-path architecture that introduces a global memory path in addition to a CNN path, allowing direct communication with any CNN layers. As a result, MaX-DeepLab shows a significant 7.1% PQ gain in the box-free regime on the challenging COCO dataset, closing the gap between box-based and box-free methods for the first time. A small variant of MaX-DeepLab improves 3.0% PQ over DETR with similar parameters and M-Adds. Furthermore, MaX-DeepLab, without test time augmentation, achieves new state-of-the-art 51.3% PQ on COCO test-dev set.

Pose-guided person image generation usually involves using paired source-target images to supervise the training, which significantly increases the data preparation effort and limits the application of the models. To deal with this problem, we propose a novel multi-level statistics transfer model, which disentangles and transfers multi-level appearance features from person images and merges them with pose features to reconstruct the source person images themselves. So that the source images can be used as supervision for self-driven person image generation. Specifically, our model extracts multi-level features from the appearance encoder and learns the optimal appearance representation through attention mechanism and attributes statistics. Then we transfer them to a pose-guided generator for re-fusion of appearance and pose. Our approach allows for flexible manipulation of person appearance and pose properties to perform pose transfer and clothes style transfer tasks. Experimental results on the DeepFashion dataset demonstrate our method's superiority compared with state-of-the-art supervised and unsupervised methods. In addition, our approach also performs well in the wild.

High-quality high-resolution (HR) magnetic resonance (MR) images afford more detailed information for reliable diagnosis and quantitative image analyses. Deep convolutional neural networks (CNNs) have shown promising ability for MR image super-resolution (SR) given low-resolution (LR) MR images. The LR MR images usually share some visual characteristics: repeating patterns, relatively simpler structures, and less informative background. Most previous CNN-based SR methods treat the spatial pixels (including the background) equally. They also fail to sense the entire space of the input, which is critical for high-quality MR image SR. To address those problems, we propose squeeze and excitation reasoning attention networks (SERAN) for accurate MR image SR. We propose to squeeze attention from global spatial information of the input and obtain global descriptors. Such global descriptors enhance the network's ability to focus on more informative regions and structures in MR images. We further build relationship among those global descriptors and propose primitive relationship reasoning attention. The global descriptors are further refined with the learned attention. To fully make use of the aggregated information, we adaptively recalibrate feature responses with learned adaptive attention vectors. These attention vectors select a subset of global descriptors to complement each spatial location for accurate details and texture reconstruction. We propose squeeze and excitation attention with residual scaling, which not only stabilizes the training but also makes it flexible to other basic networks. Extensive experiments show the effectiveness of our proposed SERAN, which clearly surpasses state-of-the-art methods on benchmarks quantitatively and qualitatively.

High-definition maps (HD maps) are a key component of most modern self-driving systems due to their valuable semantic and geometric information. Unfortunately, building HD maps has proven hard to scale due to their cost as well as the requirements they impose in the localization system that has to work everywhere with centimeter-level accuracy. Being able to drive without an HD map would be very beneficial to scale self-driving solutions as well as to increase the failure tolerance of existing ones (e.g., if localization fails or the map is not up-to-date). Towards this goal, we propose an end-to-end approach to mapless driving where the input is raw sensor data and a high-level command (e.g., turn left at the intersection). We then predict intermediate representations in the form of an online map and the current and future state of dynamic agents, and exploit them in a novel neural motion planner to make interpretable decisions taking into account uncertainty. We show that our approach is significantly safer, more comfortable, and can follow commands better than the baselines in challenging long-term closed-loop simulations, as well as when compared to an expert driver in a large-scale real-world dataset.

Over the past few years, the field of scene text detection has progressed rapidly that modern text detectors are able to hunt text in various challenging scenarios. However, they might still fall short when handling text instances of extreme aspect ratios and varying scales. To tackle such difficulties, we propose in this paper a new algorithm for scene text detection, which puts forward a set of strategies to significantly improve the quality of text localization. Specifically, a Text Feature Alignment Module (TFAM) is proposed to dynamically adjust the receptive fields of features based on initial raw detections; a Position-Aware Non-Maximum Suppression (PA-NMS) module is devised to selectively concentrate on reliable raw detections and exclude unreliable ones; besides, we propose an Instance-wise IoU loss for balanced training to deal with text instances of different scales. An extensive ablation study demonstrates the effectiveness and superiority of the proposed strategies. The resulting text detection system, which integrates the proposed strategies with a leading scene text detector EAST, achieves state-of-the-art or competitive performance on various standard benchmarks for text detection while keeping a fast running speed.

Detecting out-of-distribution (OOD) inputs is a central challenge for safely deploying machine learning models in the real world. Existing solutions are mainly driven by small datasets, with low resolution and very few class labels (e.g., CIFAR). As a result, OOD detection for large-scale image classification tasks remains largely unexplored. In this paper, we bridge this critical gap by proposing a group-based OOD detection framework, along with a novel OOD scoring function termed MOS. Our key idea is to decompose the large semantic space into smaller groups with similar concepts, which allows simplifying the decision boundaries between in- vs. out-of-distribution data for effective OOD detection. Our method scales substantially better for high-dimensional class space than previous approaches. We evaluate models trained on ImageNet against four carefully curated OOD datasets, spanning diverse semantics. MOS establishes state-of-the-art performance, reducing the average FPR95 by 14.33% while achieving 6x speedup in inference compared to the previous best method.

Out-of-distribution (OOD) detection is essential to prevent anomalous inputs from causing a model to fail during deployment. While improved OOD detection methods have emerged, they often rely on the final layer outputs and require a full feedforward pass for any given input. In this paper, we propose a novel framework, multi-level out-of-distribution detection MOOD, which exploits intermediate classifier outputs for dynamic and efficient OOD inference. We explore and establish a direct relationship between the OOD data complexity and optimal exit level, and show that easy OOD examples can be effectively detected early without propagating to deeper layers. At each exit, the OOD examples can be distinguished through our proposed adjusted energy score, which is both empirically and theoretically suitable for networks with multiple classifiers. We extensively evaluate MOOD across 10 OOD datasets spanning a wide range of complexities. Experiments demonstrate that MOOD achieves up to 71.05% computational reduction in inference, while maintaining competitive OOD detection performance.

We propose a deep network that can be trained to tackle image reconstruction and classification problems that involve detection of multiple object instances, without any supervision regarding their whereabouts. The network learns to extract the most significant top-K patches, and feeds these patches to a task-specific network -- e.g., auto-encoder or classifier -- to solve a domain specific problem. The challenge in training such a network is the non-differentiable top-K selection process. To address this issue, we lift the training optimization problem by treating the result of top-K selection as a slack variable, resulting in a simple, yet effective, multi-stage training. Our method is able to learn to detect recurrent structures in the training dataset by learning to reconstruct images. It can also learn to localize structures when only knowledge on the occurrence of the object is provided, and in doing so it outperforms the state-of-the-art.

Weakly supervised video anomaly detection (WS-VAD) is to distinguish anomalies from normal events based on discriminative representations. Most existing works are limited in insufficient video representations. In this work, we develop a multiple instance self-training framework (MIST) to efficiently refine task-specific discriminative representations with only video-level annotations. In particular, MIST is composed of 1) a multiple instance pseudo label generator, which adapts a sparse continuous sampling strategy to produce more reliable clip-level pseudo labels, and 2) a self-guided attention boosted feature encoder that aims to automatically focus on anomalous regions in frames while extracting task-specific representations. Moreover, we adopt a self-training scheme to optimize both components and finally obtain a task-specific feature encoder. Extensive experiments on two public datasets demonstrate the efficacy of our method, and our method performs comparably or even better with existing supervised and weakly supervised methods, specifically obtaining a frame-level AUC 94.83% on ShanghaiTech.

High quality Machine Learning (ML) models are often considered valuable intellectual property by companies. Model Stealing (MS) attacks allow an adversary with black-box access to a ML model to replicate its functionality by training a clone model using the predictions of the target model for different inputs. However, best available existing MS attacks fail to produce a high-accuracy clone without access to the target dataset or a representative dataset necessary to query the target model. In this paper, we show that preventing access to the target dataset is not an adequate defense to protect a model. We propose MAZE -- a data-free model stealing attack using zeroth-order gradient estimation that produces high-accuracy clones. In contrast to prior works, MAZE uses only synthetic data created using a generative model to perform MS. Our evaluation with four image classification models shows that MAZE provides a normalized clone accuracy in the range of 0.90x to 0.99x, and outperforms even the recent attacks that rely on partial data (JBDA, clone accuracy 0.13x to 0.69x) and on surrogate data (KnockoffNets, clone accuracy 0.52x to 0.97x). We also study an extension of MAZE in the partial-data setting and develop MAZE-PD, which generates synthetic data closer to the target distribution. MAZE-PD further improves the clone accuracy 0.97x to 1.0x) and reduces the query budget required for the attack by 2x-24x.