CVPR 2021 Papers

Neural signed distance functions (SDFs) are emerging as an effective representation for 3D shapes. State-of-the-art methods typically encode the SDF with a large, fixed-size neural network to approximate complex shapes with implicit surfaces. Rendering with these large networks is, however, computationally expensive since it requires many forward passes through the network for every pixel, making these representations impractical for real-time graphics. We introduce an efficient neural representation that, for the first time, enables real-time rendering of high-fidelity neural SDFs, while achieving state-of-the-art geometry reconstruction quality. We represent implicit surfaces using an octree-based feature volume which adaptively fits shapes with multiple discrete levels of detail (LODs), and enables continuous LOD with SDF interpolation. We further develop an efficient algorithm to directly render our novelneural SDF representation in real-time by querying only the necessary LODswith sparse octree traversal. We show that our representation is 2-3 orders of magnitude more efficient in terms of rendering speed compared to previous works. Furthermore, it produces state-of-the-art reconstruction quality for complex shapes under both 3D geometric and 2D image-space metrics.

RGB-Infrared person re-identification (RGB-IR ReID) is a challenging cross-modality retrieval problem, which aims at matching the person-of-interest over visible and infrared camera views. Most existing works achieve performance gains through manually-designed feature selection modules, which often require significant domain knowledge and rich experience. In this paper, we study a general paradigm, termed Neural Feature Search (NFS), to automate the process of feature selection. Specifically, NFS combines a dual-level feature search space and a differentiable search strategy to jointly select identity-related cues in coarse-grained channels and fine-grained spatial pixels. This combination allows NFS to adaptively filter background noises and concentrate on informative parts of human bodies in a data-driven manner. Moreover, a cross-modality contrastive optimization scheme further guides NFS to search features that can minimize modality discrepancy whilst maximizing inter-class distance. Extensive experiments on mainstream benchmarks demonstrate that our method outperforms state-of-the-arts, especially achieving better performance on the RegDB dataset with significant improvement of 11.20% and 8.64% in Rank-1 and mAP, respectively.

We present deep neural network methodology to reconstruct the 3d pose and shape of people, including hand gestures and facial expression, given an input RGB image. We rely on a recently introduced, expressive full body statistical 3d human model, GHUM, trained end-to-end, and learn to reconstruct its pose and shape state in a self-supervised regime. Central to our methodology, is a learning to learn and optimize approach, referred to as HUman Neural Descent (HUND), which avoids both second-order differentiation when training the model parameters, and expensive state gradient descen tin order to accurately minimize a semantic differentiable rendering loss at test time. Instead, we rely on novel recurrent stages to update the pose and shape parameters such that not only losses are minimized effectively, but the process is meta-regularized in order to ensure end-progress. HUND's symmetry between training and testing makes it the first 3d human sensing architecture to natively support different operating regimes including self-supervised ones. In diverse tests, we show that HUND achieves very competitive results in datasets like H3.6M and 3DPW, as well as good quality 3d reconstructions for complex imagery collected in-the-wild.

We introduce Neural Deformation Graphs for globally-consistent deformation tracking and 3D reconstruction of non-rigid objects. Specifically, we implicitly model a deformation graph via a deep neural network. This neural deformation graph does not rely on any object-specific structure and, thus, can be applied to general non-rigid deformation tracking. Our method globally optimizes this neural graph on a given sequence of depth camera observations of a non-rigidly moving object. Based on explicit viewpoint consistency as well as inter-frame graph and surface consistency constraints, the underlying network is trained in a self-supervised fashion. We additionally optimize for the geometry of the object with an implicit deformable multi-MLP shape representation. Our approach does not assume sequential input data, thus enabling robust tracking of fast motions or even temporally disconnected recordings. Our experiments demonstrate that our Neural Deformation Graphs outperform state-of-the-art non-rigid reconstruction approaches both qualitatively and quantitatively, with 64% improved reconstruction and 54% improved deformation tracking performance. Code is publicly available.

Very recently, the Neural Cellular Automata (NCA) has been proposed to simulate the morphogenesis process with deep networks. NCA learns to grow an image starting from a fixed single pixel. In this work, we show that the neural network (NN) architecture of the NCA can be encapsulated in a larger NN. This allows us to propose a new model that encodes a manifold of NCA, each of them capable of generating a distinct image. Therefore, we are effectively learning an embedding space of CA, which shows generalization capabilities. We accomplish this by introducing dynamic convolutions inside an Auto-Encoder architecture, for the first time used to join two different sources of information, the encoding and cell's environment information. In biological terms, our approach would play the role of the transcription factors, modulating the mapping of genes into specific proteins that drive cellular differentiation, which occurs right before the morphogenesis. We thoroughly evaluate our approach in a dataset of synthetic emojis and also in real images of CIFAR-10. Our model introduces a general-purpose network, which can be used in a broad range of problems beyond image generation.

We present a controllable camera simulator based on deep neural networks to synthesize raw image data under different camera settings, including exposure time, ISO, and aperture. The proposed simulator includes an exposure module that utilizes the principle of modern lens designs for correcting the luminance level. It also contains a noise module using the noise level function and an aperture module with adaptive attention to simulate the side effects on noise and defocus blur. To facilitate the learning of a simulator model, we collect a dataset of the 10,000 raw images of 450 scenes with different exposure settings. Quantitative experiments and qualitative comparisons show that our approach outperforms relevant baselines in raw data synthesize on multiple cameras. Furthermore, the camera simulator enables various applications, including large-aperture enhancement, HDR, auto exposure, and data augmentation for training local feature detectors. Our work represents the first attempt to simulate a camera sensor's behavior leveraging both the advantage of traditional raw sensor features and the power of data-driven deep learning.

This paper addresses the challenge of novel view synthesis for a human performer from a very sparse set of camera views. Some recent works have shown that learning implicit neural representations of 3D scenes achieves remarkable view synthesis quality given dense input views. However, the representation learning will be ill-posed if the views are highly sparse. To solve this ill-posed problem, our key idea is to integrate observations over video frames. To this end, we propose Neural Body, a new human body representation which assumes that the learned neural representations at different frames share the same set of latent codes anchored to a deformable mesh, so that the observations across frames can be naturally integrated. The deformable mesh also provides geometric guidance for the network to learn 3D representations more efficiently. Experiments on a newly collected multi-view dataset show that our approach outperforms prior works by a large margin in terms of the novel view synthesis quality. We also demonstrate the capability of our approach to reconstruct a moving person from a monocular video on the People-Snapshot dataset. We will release the code and dataset for reproducibility.

Real-world scenes have a dynamic range of up to 280 dB that today's imaging sensors cannot directly capture. Existing live vision pipelines tackle this fundamental challenge by relying on high dynamic range (HDR) sensors that try to recover HDR images from multiple captures with different exposures. While HDR sensors substantially increase the dynamic range, they are not without disadvantages, including severe artifacts for dynamic scenes, reduced fill-factor, lower resolution, and high sensor cost. At the same time, traditional auto-exposure methods for low-dynamic range sensors have advanced as proprietary methods relying on image statistics separated from downstream vision algorithms. In this work, we revisit auto-exposure control as an alternative to HDR sensors. We propose a neural network for exposure selection that is trained jointly, end-to-end with an object detector and an image signal processing (ISP) pipeline. To this end, we use an HDR dataset for automotive object detection and an HDR training procedure. We validate that the proposed neural auto-exposure control, which is tailored to object detection, outperforms conventional auto-exposure methods by more than 6 points in mean average precision (mAP).

In this paper, we investigate a new variant of neural architecture search (NAS) paradigm -- searching with random labels (RLNAS). The task sounds counter-intuitive for most existing NAS algorithms since random label provides few information on the performance of each candidate architecture. Instead, we propose a novel NAS framework based on ease-of-convergence hypothesis, which requires only random labels during searching. The algorithm involves two steps: first, we train a SuperNet using random labels; second, from the SuperNet we extract the sub-network whose weights change most significantly during the training. Extensive experiments are evaluated on multiple datasets (e.g. NAS-Bench-201 and ImageNet) and multiple search spaces (e.g. DARTS-like and MobileNet-like). Very surprisingly, RLNAS achieves comparable or even better results compared with state-of-the-art NAS methods such as PC-DARTS, Single Path One-Shot, even though the counterparts utilize full ground truth labels for searching. We hope our finding could inspire new understandings on the essential of NAS.

Recent work has demonstrated that volumetric scene representations combined with differentiable volume rendering can enable photo-realistic rendering for challenging scenes that mesh reconstruction fails on. However, these methods entangle geometry and appearance in a ""black-box"" volume that cannot be edited. Instead, we present an approach that explicitly disentangles geometry--represented as a continuous 3D volume--from appearance--represented as a continuous 2D texture map. We achieve this by introducing a 3D-to-2D texture mapping (or surface parameterization) network into volumetric representations. We constrain this texture mapping network using an additional 2D-to-3D inverse mapping network and a novel cycle consistency loss to make 3D surface points map to 2D texture points that map back to the original 3D points. We demonstrate that this representation can be reconstructed using only multi-view image supervision and generates high-quality rendering results. More importantly,by separating geometry and texture, we allow users to edit appearance by simply editing 2D texture maps.

Network quantization aims at reducing bit-widths of weights and/or activations, particularly important for implementing deep neural networks with limited hardware resources. Most methods use the straight-through estimator (STE) to train quantized networks, which avoids a zero-gradient problem by replacing a derivative of a discretizer (i.e., a round function) with that of an identity function. Although quantized networks exploiting the STE have shown decent performance, the STE is sub-optimal in that it simply propagates the same gradient without considering discretization errors between inputs and outputs of the discretizer. In this paper, we propose an element-wise gradient scaling (EWGS), a simple yet effective alternative to the STE, training a quantized network better than the STE in terms of stability and accuracy. Given a gradient of the discretizer output, EWGS adaptively scales up or down each gradient element, and uses the scaled gradient as the one for the discretizer input to train quantized networks via backpropagation. The scaling is performed depending on both the sign of each gradient element and an error between the continuous input and discrete output of the discretizer. We adjust a scaling factor adaptively using Hessian information of a network. We show extensive experimental results on the image classification datasets, including CIFAR-10 and ImageNet, with diverse network architectures under a wide range of bit-width settings, demonstrating the effectiveness of our method.

Channel pruning is a class of powerful methods for model compression. When pruning a neural network, it's ideal to obtain a sub-network with higher accuracy. However, a sub-network does not necessarily have high accuracy with low classification loss (loss-metric mismatch). In the paper, we first consider the loss-metric mismatch problem for pruning and propose a novel channel pruning method for Convolutional Neural Networks (CNNs) by directly maximizing the performance (i.e., accuracy) of sub-networks. Specifically, we train a stand-alone neural network to predict sub-networks' performance and then maximize the output of the network as a proxy of accuracy to guide pruning. Training such a performance prediction network efficiently is not an easy task, and it may potentially suffer from the problem of catastrophic forgetting and the imbalance distribution of sub-networks. To deal with this challenge, we introduce a corresponding episodic memory to update and collect sub-networks during the pruning process. In the experiment section, we further demonstrate that the gradients from the performance prediction network and the classification loss have different directions. Extensive experimental results show that the proposed method can achieve state-of-the-art performance with ResNet, MobileNetV2, and ShuffleNetV2+ on ImageNet and CIFAR-10.

Neural architecture search (NAS) typically consists of three main steps: training a super-network, training and evaluating sampled deep neural networks (DNNs), and training the discovered DNN. Most of the existing efforts speed up some steps at the cost of a significant slowdown of other steps or sacrificing the support of non-differentiable search metrics. The unbalanced reduction in the time spent per step limits the total search time reduction, and the inability to support non-differentiable search metrics limits the performance of discovered DNNs. In this paper, we present NetAdaptV2 with three innovations to better balance the time spent for each step while supporting non-differentiable search metrics. First, we propose channel-level bypass connections that merge network depth and layer width into a single search dimension to reduce the time for training and evaluating sampled DNNs. Second, ordered dropout is proposed to train multiple DNNs in a single forward-backward pass to decrease the time for training a super-network. Third, we propose the multi-layer coordinate descent optimizer that considers the interplay of multiple layers in each iteration of optimization to improve the performance of discovered DNNs while supporting non-differentiable search metrics. With these innovations, NetAdaptV2 reduces the total search time by up to 5.8x on ImageNet and 2.4x on NYU Depth V2, respectively, and discovers DNNs with better accuracy-latency/accuracy-MAC trade-offs than state-of-the-art NAS works. Moreover, the discovered DNN outperforms NAS-discovered MobileNetV3 by 1.8% higher top-1 accuracy with the same latency.

Extracting geometric features from 3D models is a common first step in applications such as 3D registration, tracking, and scene flow estimation. Many hand-crafted and learning-based methods aim to produce consistent and distinguishable geometric features for 3D models with partial overlap. These methods work well in cases where the point density and scale of the overlapping 3D objects are similar, but struggle in applications where 3D data are obtained independently with unknown global scale and scene overlap. Unfortunately, instances of this resolution mismatch are common in practice, e.g., when aligning data from multiple sensors. In this work, we introduce a new normalization technique, Batch-Neighborhood Normalization, aiming to improve robustness to mean-std variation of local feature distributions that presumably can happen in samples with varying point density. We empirically demonstrate that the presented normalization method's performance compares favorably to comparison methods in indoor and outdoor environments, and on a clinical dataset, on common point registration benchmarks in both standard and, particularly, resolution-mismatch settings. The source code and clinical dataset are available at https://github.com/lppllppl920/NeighborhoodNormalization-Pytorch.

In this paper, we address Novel Class Discovery (NCD), the task of unveiling new classes in a set of unlabeled samples given a labeled dataset with known classes. We exploit the peculiarities of NCD to build a new framework, named Neighborhood Contrastive Learning (NCL), to learn discriminative representations that are important to clustering performance. Our contribution is twofold. First, we find that a feature extractor trained on the labeled set generates representations in which a generic query sample and its neighbors are likely to share the same class. We exploit this observation to retrieve and aggregate pseudo positive pairs with contrastive learning, thus encouraging the model to learn more discriminative representations. Second, we notice that most of the instances are easily discriminated by the network, contributing less to the contrastive loss. To overcome this issue, we propose to generate hard negatives by mixing labeled and unlabeled samples in the feature space. We experimentally demonstrate that these two ingredients significantly contribute to clustering performance and lead our model to outperform state of the art by a large margin (e.g., clustering accuracy +13% on CIFAR-100 and +8% on ImageNet).

In the last few years, image denoising has benefited a lot from the fast development of neural networks. However, the requirement of large amounts of noisy-clean image pairs for supervision limits the wide use of these models. Although there have been a few attempts in training an image denoising model with only single noisy images, existing self-supervised denoising approaches suffer from inefficient network training, loss of useful information, or dependence on noise modeling. In this paper, we present a very simple yet effective method named Neighbor2Neighbor to train an effective image denoising model with only noisy images. Firstly, a random neighbor sub-sampler is proposed for the generation of training image pairs. In detail, input and target used to train a network are images sub-sampled from the same noisy image, satisfying the requirement that paired pixels of paired images are neighbors and have very similar appearance with each other. Secondly, a denoising network is trained on sub-sampled training pairs generated in the first stage, with a proposed regularizer as additional loss for better performance. The proposed Neighbor2Neighbor framework is able to enjoy the progress of state-of-the-art supervised denoising networks in network architecture design. Moreover, it avoids heavy dependence on the assumption of the noise distribution. We explain our approach from a theoretical perspective and further validate it through extensive experiments, including synthetic experiments with different noise distributions in sRGB space and real-world experiments on a denoising benchmark dataset in raw-RGB space.

Deep clustering gradually becomes an important branch in unsupervised learning methods. However, current approaches hardly take into consideration the semantic sample relationships that existed in both local and global features. In addition, since the deep features are updated on-the-fly, relying on these sample relationships may construct more semantically confident sample pairs, leading to inferior performance. To tackle this issue, we propose a method called Nearest Neighbor Matching (NNM) to match samples with their nearest neighbors from both local (batch) and global (overall) levels. Specifically, for the local level, we match the nearest neighbors based on batch embedded features, as for the global one, we match neighbors from overall embedded features. To keep the clustering assignment consistent in both neighbors and classes, we frame consistent loss and class contrastive loss for both local and global levels. Experimental results on three benchmark datasets demonstrate the superiority of our new model against state-of-the-art methods. Particularly on the STL-10 dataset, our method can achieve supervised performance. As for the CIFAR-100 dataset, our NNM leads 3.7% against the latest comparison method. Our code will be available at https://github.com/ZhiyuanDang/NNM.

We present NeX, a new approach to novel view synthesis based on enhancements of multiplane image (MPI) that can reproduce next-level view-dependent effects--in real time. Unlike traditional MPI that uses a set of simple RGBa planes, our technique models view-dependent effects by instead parameterizing each pixel as a linear combination of basis functions learned from a neural network. Moreover, we propose a hybrid implicit-explicit modeling strategy that improves upon fine detail and produces state-of-the-art results. Our method is evaluated on benchmark forward-facing datasets as well as our newly-introduced dataset designed to test the limit of view-dependent modeling with significantly more challenging effects such as the rainbow reflections on a CD. Our method achieves the best overall scores across all major metrics on these datasets with more than 1000x faster rendering time than the state of the art. For real-time demos, visit https://nex-mpi.github.io/

We present a method that takes as input a set of images of a scene illuminated by unconstrained known lighting, and produces as output a 3D representation that can be rendered from novel viewpoints under arbitrary lighting conditions. Our method represents the scene as a continuous volumetric function parameterized as MLPs whose inputs are a 3D location and whose outputs are the following scene properties at that input location: volume density, surface normal, material parameters, distance to the first surface intersection in any direction, and visibility of the external environment in any direction. Together, these allow us to render novel views of the object under arbitrary lighting, including indirect illumination effects. The predicted visibility and surface intersection fields are critical to our model's ability to simulate direct and indirect illumination during training, because the brute-force techniques used by prior work are intractable for lighting conditions outside of controlled setups with a single light. Our method outperforms alternative approaches for recovering relightable 3D scene representations, and performs well in complex lighting settings that have posed a significant challenge to prior work.