CVPR 2022 Papers

This paper presents a new approach to boost a single-modality (LiDAR) 3D object detector by teaching it to simulate features and responses that follow a multi-modality (LiDAR-image) detector. The approach needs LiDAR-image data only when training the single-modality detector, and once well-trained, it only needs LiDAR data at inference. We design a novel framework to realize the approach: response distillation to focus on the crucial response samples and avoid the background samples; sparse-voxel distillation to learn voxel semantics and relations from the estimated crucial voxels; a fine-grained voxel-to-point distillation to better attend to features of small and distant objects; and instance distillation to further enhance the deep-feature consistency. Experimental results on the nuScenes dataset show that our approach outperforms all SOTA LiDAR-only 3D detectors and even surpasses the baseline LiDAR-image detector on the key NDS metric, filling 72% mAP gap between the single- and multi-modality detectors.

A significant gap remains between today's visual pattern recognition models and human-level visual cognition especially when it comes to few-shot learning and compositional reasoning of novel concepts. We introduce Bongard-HOI, a new visual reasoning benchmark that focuses on compositional learning of human-object interactions (HOIs) from natural images. It is inspired by two desirable characteristics from the classical Bongard problems (BPs): 1) few-shot concept learning, and 2) context-dependent reasoning. We carefully curate the few-shot instances with hard negatives, where positive and negative images only disagree on action labels, making mere recognition of object categories insufficient to complete our benchmarks. We also design multiple test sets to systematically study the generalization of visual learning models, where we vary the overlap of the HOI concepts between the training and test sets of few- shot instances, from partial to no overlaps. Bongard-HOI presents a substantial challenge to today's visual recognition models. The state-of-the-art HOI detection model achieves only 62% accuracy on few-shot binary prediction while even amateur human testers on MTurk have 91% accuracy. With the Bongard-HOI benchmark, we hope to further advance research efforts in visual reasoning, especially in holistic perception-reasoning systems and better representation learning.

We propose BokehMe, a hybrid bokeh rendering framework that marries a neural renderer with a classical physically motivated renderer. Given a single image and a potentially imperfect disparity map, BokehMe generates high-resolution photo-realistic bokeh effects with adjustable blur size, focal plane, and aperture shape. To this end, we analyze the errors from the classical scattering-based method and derive a formulation to calculate an error map. Based on this formulation, we implement the classical renderer by a scattering-based method and propose a two-stage neural renderer to fix the erroneous areas from the classical renderer. The neural renderer employs a dynamic multi-scale scheme to efficiently handle arbitrary blur sizes, and it is trained to handle imperfect disparity input. Experiments show that our method compares favorably against previous methods on both synthetic image data and real image data with predicted disparity. A user study is further conducted to validate the advantage of our method.

We present Block-NeRF, a variant of Neural Radiance Fields that can represent large-scale environments. Specifically, we demonstrate that when scaling NeRF to render city-scale scenes spanning multiple blocks, it is vital to decompose the scene into individually trained NeRFs. This decomposition decouples rendering time from scene size, enables rendering to scale to arbitrarily large environments, and allows per-block updates of the environment. We adopt several architectural changes to make NeRF robust to data captured over months under different environmental conditions. We add appearance embeddings, learned pose refinement, and controllable exposure to each individual NeRF, and introduce a procedure for aligning appearance between adjacent NeRFs so that they can be seamlessly combined. We build a grid of Block-NeRFs from 2.8 million images to create the largest neural scene representation to date, capable of rendering an entire neighborhood of San Francisco.

Shadow removal, which aims to restore the background in the shadow regions, is challenging due to the highly ill-posed nature. Most existing deep learning-based methods individually remove the shadow by only considering the content of the matched paired images, barely taking into account the auxiliary supervision of shadow generation on shadow removal procedure. In this work, we argue that shadow removal and generation are interrelated and could provide useful informative supervision for each other. Specifically, we propose a new Bijective Mapping Network (BMNet), which couples the learning procedures of shadow removal and shadow generation in a unified parameter-shared framework. With consistent two-way constraints and synchronous optimization of the two procedures, BMNet could effectively recover the underlying background contents during the forward shadow removal procedure. In addition, through statistic analysis on real-world datasets, we observe and verify that shadow appearances under different color spectrums are inconsistent. This motivates us to design a Shadow-Invariant Color Guidance Module (SICGM), which can explicitly utilize the learned shadow-invariant color information to guide network color restoration, thereby further reducing color-bias effects. Experiments on the representative ISTD, ISTD+ and SRD benchmarks show that our proposed network outperforms the state-of-the-art method [??] in de-shadowing performance, while only using its 0.25% network parameters and 6.25% floating point operations (FLOPs).

3D single object tracking (3D SOT) in LiDAR point clouds plays a crucial role in autonomous driving. Current approaches all follow the Siamese paradigm based on appearance matching. However, LiDAR point clouds are usually textureless and incomplete, which hinders effective appearance matching. Besides, previous methods greatly overlook the critical motion clues among targets. In this work, beyond 3D Siamese tracking, we introduce a motion-centric paradigm to handle 3D SOT from a new perspective. Following this paradigm, we propose a matching-free two-stage tracker M^2-Track. At the 1^st-stage, M^2-Track localizes the target within successive frames via motion transformation. Then it refines the target box through motion-assisted shape completion at the 2^nd-stage. Extensive experiments confirm that M^2-Track significantly outperforms previous state-of-the-arts on three large-scale datasets while running at 57FPS ( 8%, 17%, and 22%) precision gains on KITTI, NuScenes, and Waymo Open Dataset respectively). Further analysis verifies each component's effectiveness and shows the motion-centric paradigm's promising potential when combined with appearance matching. Code will be made available at https://github.com/Ghostish/Open3DSOT.

Backdoor attacks aim to cause misclassification of a subject model by stamping a trigger to inputs. Backdoors could be injected through malicious training and naturally exist. Deriving backdoor trigger for a subject model is critical to both attack and defense. A popular trigger inversion method is by optimization. Existing methods are based on finding a smallest trigger that can uniformly flip a set of input samples by minimizing a mask. The mask defines the set of pixels that ought to be perturbed. We develop a new optimization method that directly minimizes individual pixel changes, without using a mask. Our experiments show that compared to existing methods, the new one can generate triggers that require a smaller number of input pixels to be perturbed, have a higher attack success rate, and are more robust. They are hence more desirable when used in real-world attacks and more effective when used in defense. Our method is also more cost-effective.

Generalization under distributional shift is an open challenge for machine learning. Invariant Risk Minimization (IRM) is a promising framework to tackle this issue by extracting invariant features. However, despite the potential and popularity of IRM, recent works have reported negative results of it on deep models. We argue that the failure can be primarily attributed to deep models' tendency to overfit the data. Specifically, our theoretical analysis shows that IRM degenerates to empirical risk minimization (ERM) when overfitting occurs. Our empirical evidence also provides supports: IRM methods that work well in typical settings significantly deteriorate even if we slightly enlarge the model size or lessen the training data. To alleviate this issue, we propose Bayesian Invariant Risk Minimization (BIRM) by introducing Bayesian inference into the IRM. The key motivation is to estimate the penalty of IRM based on the posterior distribution of classifiers (as opposed to a single classifier), which is much less prone to overfitting. Extensive experimental results on four datasets demonstrate that BIRM consistently outperforms the existing IRM baselines significantly.

Graph neural networks (GNNs) have achieved state-ofthe-art performance in many graph-based tasks such as node classification and graph classification. However, many recent works have demonstrated that an attacker can mislead GNN models by slightly perturbing the graph structure. Existing attacks to GNNs are either under the less practical threat model where the attacker is assumed to access the GNN model parameters, or under the practical black-box threat model but consider perturbing node features that are shown to be not enough effective. In this paper, we aim to bridge this gap and consider black-box attacks to GNNs with structure perturbation as well as with theoretical guarantees. We propose to address this challenge through bandit techniques. Specifically, we formulate our attack as an online optimization with bandit feedback. This original problem is essentially NP-hard due to the fact that perturbing the graph structure is a binary optimization problem. We then propose an online attack based on bandit optimization which is proven to be sublinear to the query number T, i.e., O(N^ 1/2 T^ 3/4 ) where N is the number of nodes in the graph. Finally, we evaluate our proposed attack by conducting experiments over multiple datasets and GNN models. The experimental results on various citation graphs and image graphs show that our attack is both effective and efficient.

Audio-visual learning helps to comprehensively understand the world, by integrating different senses. Accordingly, multiple input modalities are expected to boost model performance, but we actually find that they are not fully exploited even when the multi-modal model outperforms its uni-modal counterpart. Specifically, in this paper we point out that existing audio-visual discriminative models, in which uniform objective is designed for all modalities, could remain under-optimized uni-modal representations, caused by another dominated modality in some scenarios, e.g., sound in blowing wind event, vision in drawing picture event, etc. To alleviate this optimization imbalance, we propose on-the-fly gradient modulation to adaptively control the optimization of each modality, via monitoring the discrepancy of their contribution towards the learning objective. Further, an extra Gaussian noise that changes dynamically is introduced to avoid possible generalization drop caused by gradient modulation. As a result, we achieve considerable improvement over common fusion methods on different audio-visual tasks, and this simple strategy can also boost existing multi-modal methods, which illustrates its efficacy and versatility.

Data imbalance exists ubiquitously in real-world visual regressions, e.g., age estimation and pose estimation, hurting the model's generalizability and fairness. Thus, imbalanced regression gains increasing research attention recently. Compared to imbalanced classification, imbalanced regression focuses on continuous labels, which can be boundless and high-dimensional and hence more challenging. In this work, we identify that the widely used Mean Square Error (MSE) loss function can be ineffective in imbalanced regression. We revisit MSE from a statistical view and propose a novel loss function, Balanced MSE, to accommodate the imbalanced training label distribution. We further design multiple implementations of Balanced MSE to tackle different real-world scenarios, particularly including the one that requires no prior knowledge about the training label distribution. Moreover, to the best of our knowledge, Balanced MSE is the first general solution to high-dimensional imbalanced regression. Extensive experiments on both synthetic and three real-world benchmarks demonstrate the effectiveness of Balanced MSE.

Driving 3D characters to dance following a piece of music is highly challenging due to the spatial constraints applied to poses by choreography norms. In addition, the generated dance sequence also needs to maintain temporal coherency with different music genres. To tackle these challenges, we propose a novel music-to-dance framework, Bailando, with two powerful components: 1) a choreographic memory that learns to summarize meaningful dancing units from 3D pose sequence to a quantized codebook, 2) an actor-critic Generative Pre-trained Transformer (GPT) that composes these units to a fluent dance coherent to the music. With the learned choreographic memory, dance generation is realized on the quantized units that meet high choreography standards, such that the generated dancing sequences are confined within the spatial constraints. To achieve synchronized alignment between diverse motion tempos and music beats, we introduce an actor-critic-based reinforcement learning scheme to the GPT with a newly-designed beat-align reward function. Extensive experiments on the standard benchmark demonstrate that our proposed framework achieves state-of-the-art performance both qualitatively and quantitatively. Notably, the learned choreographic memory is shown to discover human-interpretable dancing-style poses in an unsupervised manner.

Large-scale unlabeled data has spurred recent progress in self-supervised learning methods that learn rich visual representations. State-of-the-art self-supervised methods for learning representations from images (e.g., MoCo, BYOL, MSF) use an inductive bias that random augmentations (e.g., random crops) of an image should produce similar embeddings. We show that such methods are vulnerable to backdoor attacks -- where an attacker poisons a small part of the unlabeled data by adding a trigger (image patch chosen by the attacker) to the images. The model performance is good on clean test images, but the attacker can manipulate the decision of the model by showing the trigger at test time. Backdoor attacks have been studied extensively in supervised learning and to the best of our knowledge, we are the first to study them for self-supervised learning. Backdoor attacks are more practical in self-supervised learning, since the use of large unlabeled data makes data inspection to remove poisons prohibitive. We show that in our targeted attack, the attacker can produce many false positives for the target category by using the trigger at test time. We also propose a defense method based on knowledge distillation that succeeds in neutralizing the attack. Our code is available here: https://github.com/UMBCvision/SSL-Backdoor

As a prerequisite of many text-related tasks such as text erasing and text style transfer, text segmentation arouses more and more attention recently. Current researches mainly focus on only English characters and digits, while few work studies Chinese characters due to the lack of public large-scale and high-quality Chinese datasets, which limits the practical application scenarios of text segmentation. Different from English which has a limited alphabet of letters, Chinese has much more basic characters with complex structures, making the problem more difficult to deal with. To better analyze this problem, we propose the Bi-lingual Text Segmentation (BTS) dataset, a benchmark that covers various common Chinese scenes including 14,250 diverse and fine-annotated text images. BTS mainly focuses on Chinese characters, and also contains English words and digits. We also introduce Prior Guided Text Segmentation Network (PGTSNet), the first baseline to handle bi-lingual and complex-structured text segmentation. A plug-in text region highlighting module and a text perceptual discriminator are proposed in PGTSNet to supervise the model with text prior, and guide for more stable and finer text segmentation. A variation loss is also employed for suppressing background noise under complex scene. Extensive experiments are conducted not only to demonstrate the necessity and superiority of the proposed dataset BTS, but also to show the effectiveness of the proposed PGTSNet compared with a variety of state-of-the-art text segmentation methods.

Prior work for articulated 3D shape reconstruction often relies on specialized multi-view and depth sensors or pre-built deformable 3D models. Such methods do not scale to diverse sets of objects in the wild. We present a method that requires neither of them. It builds high-fidelity, articulated 3D models from many monocular casual videos in a differentiable rendering framework. Our key insight is to merge three schools of thought: (1) classic deformable shape models that make use of articulated bones and blend skinning, (2) canonical embeddings that establish correspondences between pixels and a canonical 3D model, and (3) volumetric neural radiance fields (NeRFs) that are amenable to gradient-based optimization. We introduce neural blend skinning models that allow for differentiable and invertible articulated deformations. When combined with canonical embeddings, such models allow us to establish dense correspondences across videos that can be self-supervised with cycle consistency. On real and synthetic datasets, our method shows higher-fidelity 3D reconstructions than prior works for humans and animals, with the ability to render realistic images from novel viewpoints.

Coordinate-based networks have emerged as a powerful tool for 3D representation and scene reconstruction. These networks are trained to map continuous input coordinates to the value of a signal at each point. Still, current architectures are black boxes: their spectral characteristics cannot be easily analyzed, and their behavior at unsupervised points is difficult to predict. Moreover, these networks are typically trained to represent a signal at a single scale, so naive downsampling or upsampling results in artifacts. We introduce band-limited coordinate networks (BACON), a network architecture with an analytical Fourier spectrum. BACON has constrained behavior at unsupervised points, can be designed based on the spectral characteristics of the represented signal, and can represent signals at multiple scales without per-scale supervision. We demonstrate BACON for multiscale neural representation of images, radiance fields, and 3D scenes using signed distance functions and show that it outperforms conventional single-scale coordinate networks in terms of interpretability and quality.

Graph neural architecture search has sparked much attention as Graph Neural Networks (GNNs) have shown powerful reasoning capability in many relational tasks. However, the currently used graph search space overemphasizes learning node features and neglects mining hierarchical relational information. Moreover, due to diverse mechanisms in the message passing, the graph search space is much larger than that of CNNs. This hinders the straightforward application of classical search strategies for exploring complicated graph search space. We propose Automatic Relation-aware Graph Network Proliferation (ARGNP) for efficiently searching GNNs with a relation-guided message passing mechanism. Specifically, we first devise a novel dual relation-aware graph search space that comprises both node and relation learning operations. These operations can extract hierarchical node/relational information and provide anisotropic guidance for message passing on a graph. Second, analogous to cell proliferation, we design a network proliferation search paradigm to progressively determine the GNN architectures by iteratively performing network division and differentiation. The experiments on six datasets for four graph learning tasks demonstrate that GNNs produced by our method are superior to the current state-of-the-art hand-crafted and search-based GNNs.

Focus control (FC) is crucial for cameras to capture sharp images in challenging real-world scenarios. The autofocus (AF) facilitates the FC by automatically adjusting the focus settings. However, due to the lack of effective AF methods for the recently introduced event cameras, their FC still relies on naive AF like manual focus adjustments, leading to poor adaptation in challenging real-world conditions. In particular, the inherent differences between event and frame data in terms of sensing modality, noise, temporal resolutions, etc., bring many challenges in designing an effective AF method for event cameras. To address these challenges, we develop a novel event-based autofocus framework consisting of an event-specific focus measure called event rate (ER) and a robust search strategy called event-based golden search (EGS). To verify the performance of our method, we have collected an event-based autofocus dataset (EAD) containing well-synchronized frames, events, and focal positions in a wide variety of challenging scenes with severe lighting and motion conditions. The experiments on this dataset and additional real-world scenarios demonstrated the superiority of our method over state-of-the-art approaches in terms of efficiency and accuracy.

Since facial actions such as lip movements contain significant information about speech content, it is not surprising that audio-visual speech enhancement methods are more accurate than their audio-only counterparts. Yet, state-of-the-art approaches still struggle to generate clean, realistic speech without noise artifacts and unnatural distortions in challenging acoustic environments. In this paper, we propose a novel audio-visual speech enhancement framework for high-fidelity telecommunications in AR/VR. Our approach leverages audio-visual speech cues to generate the codes of a neural speech codec, enabling efficient synthesis of clean, realistic speech from noisy signals. Given the importance of speaker-specific cues in speech, we focus on developing personalized models that work well for individual speakers. We demonstrate the efficacy of our approach on a new audio-visual speech dataset collected in an unconstrained, large vocabulary setting, as well as existing audio-visual datasets, outperforming speech enhancement baselines on both quantitative metrics and human evaluation studies. Please see the supplemental video for qualitative results.