CVPR 2021 Papers

Semantic segmentation aims to classify every pixel of an input image. Considering the difficulty of acquiring dense labels, researchers have recently been resorting to weak labels to alleviate the annotation burden of segmentation. However, existing works mainly concentrate on expanding the seed of pseudo labels within the image's salient region. In this work, we propose a non-salient region object mining approach for weakly supervised semantic segmentation. We introduce a graph-based global reasoning unit to strengthen the classification network's ability to capture global relations among disjoint and distant regions. This helps the network activate the object features outside the salient area. To further mine the non-salient region objects, we propose to exert the segmentation network's self-correction ability. Specifically, a potential object mining module is proposed to reduce the false-negative rate in pseudo labels. Moreover, we propose a non-salient region masking module for complex images to generate masked pseudo labels. Our non-salient region masking module helps further discover the objects in the non-salient region. Extensive experiments on the PASCAL VOC dataset demonstrate state-of-the-art results compared to current methods.

The existence of noisy labels in real-world data negatively impacts the performance of deep learning models. Although much research effort has been devoted to improving robustness to noisy labels in classification tasks, the problem of noisy labels in deep metric learning (DML) remains open. In this paper, we propose a noise-resistant training technique for DML, which we name Probabilistic Ranking-based Instance Selection with Memory (PRISM). PRISM identifies noisy data in a minibatch using average similarity against image features extracted by several previous versions of the neural network. These features are stored in and retrieved from a memory bank. To alleviate the high computational cost brought by the memory bank, we introduce an acceleration method that replaces individual data points with the class centers. In extensive comparisons with 12 existing approaches under both synthetic and real-world label noise, PRISM demonstrates superior performance of up to 6.06% in Precision@1.

Removing objects from images is a challenging technical problem that is important for many applications, including mixed reality. For believable results, the shadows that the object casts should also be removed. Current inpainting-based methods only remove the object itself, leaving shadows behind, or at best require specifying shadow regions to inpaint. We introduce a deep learning pipeline for removing a shadow along with its caster. We leverage rough scene models in order to remove a wide variety of shadows (hard or soft, dark or subtle, large or thin) from surfaces with a wide variety of textures. We train our pipeline on synthetically rendered data, and show qualitative and quantitative results on both synthetic and real scenes.

Not all video frames are equally informative for recognizing an action. It is computationally infeasible to train deep networks on all video frames when actions develop over hundreds of frames. A common heuristic is uniformly sampling a small number of video frames and using these to recognize the action. Instead, here we propose full video action recognition and consider all video frames. To make this computational tractable, we first cluster all frame activations along the temporal dimension based on their similarity with respect to the classification task, and then temporally aggregate the frames in the clusters into a smaller number of representations. Our method is end-to-end trainable and computationally efficient as it relies on temporally localized clustering in combination with fast Hamming distances in feature space. We evaluate on UCF101, HMDB51, Breakfast, and Something-Something V1 and V2, where we compare favorably to existing heuristic frame sampling methods.

Most existing nighttime visibility enhancement methods focus on low light. Night images, however, do not only suffer from low light, but also from man-made light effects such as glow, glare, floodlight, etc. Hence, when the existing nighttime visibility enhancement methods are applied to these images, they intensify the effects, degrading the visibility even further. High dynamic range (HDR) imaging methods can address the low light and over-exposed regions, however they cannot remove the light effects, and thus cannot enhance the visibility in the affected regions. In this paper, given a single nighttime image as input, our goal is to enhance its visibility by increasing the dynamic range of the intensity, and thus can boost the intensity of the low light regions, and at the same time, suppress the light effects (glow, glare) simultaneously. First, we use a network to estimate the camera response function (CRF) from the input image to linearise the image. Second, we decompose the linearised image into low-frequency (LF) and high-frequency (HF) feature maps that are processed separately through two networks for light effects suppression and noise removal respectively. Third, we use a network to increase the dynamic range of the processed LF feature maps, which are then combined with the processed HF feature maps to generate the final output that has increased dynamic range and suppressed light effects. Our experiments show the effectiveness of our method in comparison with the state-of-the-art nighttime visibility enhancement methods.

Learning low-dimensional latent state space dynamics models has proven powerful for enabling vision-based planning and learning for control. We introduce a latent dynamics learning framework that is uniquely designed to induce proportional controlability in the latent space, thus enabling the use of simple and well-known PID controllers. We show that our learned dynamics model enables proportional control from pixels, dramatically simplifies and accelerates behavioural cloning of vision-based controllers, and provides interpretable goal discovery when applied to imitation learning of switching controllers from demonstration. Notably, such proportional controlability also allows for robust path following from visual demonstrations using Dynamic Movement Primitives in the learned latent space.

We present NeuroMorph, a new neural network architecture that takes as input two 3D shapes and produces in one go, i.e. in a single feed forward pass, a smooth interpolation and point-to-point correspondences between them. The interpolation, expressed as a deformation field, changes the pose of the source shape to resemble the target, but leaves the object identity unchanged. NeuroMorph uses an elegant architecture combining graph convolutions with global feature pooling to extract local features. During training, the model is incentivized to create realistic deformations by approximating geodesics on the underlying shape space manifold. This strong geometric prior allows to train our model end-to-end and in a fully unsupervised manner without requiring any manual correspondence annotations. NeuroMorph works well for a large variety of input shapes, including non-isometric pairs from different object categories. It obtains state-of-the-art results for both shape correspondence and interpolation tasks, matching or surpassing the performance of recent unsupervised and supervised methods on multiple benchmarks.

We present a novel framework named NeuralRecon for real-time 3D scene reconstruction from a monocular video. Unlike previous methods that estimate single-view depth maps separately on each key-frame and fuse them later, we propose to directly reconstruct local surfaces represented as sparse TSDF volumes for each video fragment sequentially by a neural network. A learning-based TSDF fusion module based on gated recurrent units is used to guide the network to fuse features from previous fragments. This design allows the network to capture local smoothness prior and global shape prior of 3D surfaces when sequentially reconstructing the surfaces, resulting in accurate, coherent, and real-time surface reconstruction. The experiments on ScanNet and 7-Scenes datasets show that our system outperforms state-of-the-art methods in terms of both accuracy and speed. To the best of our knowledge, this is the first learning-based system that is able to reconstruct dense coherent 3D geometry in real-time. Code is available at the project page: https://zju3dv.github.io/neuralrecon/.

4D reconstruction and rendering of human activities is critical for immersive VR/AR experience. Recent advances still fail to recover fine geometry and texture results with the level of detail present in the input images from sparse multi-view RGB cameras. In this paper, we propose NeuralHumanFVV, a real-time neural human performance capture and rendering system to generate both high-quality geometry and photo-realistic texture of human activities in arbitrary novel views. We propose a neural geometry generation scheme with a hierarchical sampling strategy for real-time implicit geometry inference, as well as a novel neural blending scheme to generate high resolution (e.g., 1k) and photo-realistic texture results in the novel views. Furthermore, we adopt neural normal blending to enhance geometry details and formulate our neural geometry and texture rendering into a multi-task learning framework. Extensive experiments demonstrate the effectiveness of our approach to achieve high-quality geometry and photo-realistic free view-point reconstruction for challenging human performances.

We present a novel online depth map fusion approach that learns depth map aggregation in a latent feature space. While previous fusion methods use an explicit scene representation like signed distance functions (SDFs), we propose a learned feature representation for the fusion. The key idea is a separation between the scene representation used for the fusion and the output scene representation, via an additional translator network. Our neural network architecture consists of two main parts: a depth and feature fusion sub-network, which is followed by a translator sub-network to produce the final surface representation (e.g. TSDF) for visualization or other tasks. Our approach is an online process, handles high noise levels, and is particularly able to deal with gross outliers common for photometric stereo-based depth maps. Experiments on real and synthetic data demonstrate improved results compared to the state of the art, especially in challenging scenarios with large amounts of noise and outliers. The source code will be made available at https://github.com/weders/NeuralFusion.

Maps are arguably one of the most fundamental concepts used to define and operate on manifold surfaces in differentiable geometry. Accordingly, in geometry processing, maps are ubiquitous and are used in many core applications, such as paramterization, shape analysis, remeshing, and deformation. Unfortunately, most computational representations of surface maps do not lend themselves to manipulation and optimization, usually entailing hard, discrete problems. While algorithms exist to solve these problems, they are problem-specific, and a general framework for surface maps is still in need. In this paper, we advocate to consider neural networks as encoding surface maps. Since neural networks can be composed on one another and are differentiable, we show it is easy to use them to define surfaces via atlases, compose them for surface-to-surface mappings, and optimize differentiable objectives relating to them, such as any notion of distortion, in a trivial manner. In our experiments, we represent surfaces by generating a neural map that approximates a UV parameterization of a 3D model. Then, we compose this map with other neural maps which we optimize with respect to distortion measures. We show that our formulation enables trivial optimization of rather elusive mapping tasks, such as maps between a collection of surfaces.

We present Neural Splines, a technique for 3D surface reconstruction that is based on random feature kernels arising from infinitely-wide shallow ReLU networks. Our method achieves state-of-the-art results, outperforming recent neural network-based techniques and widely used Poisson Surface Reconstruction (which, as we demonstrate, can also be viewed as a type of kernel method). Because our approach is based on a simple kernel formulation, it is easy to analyze and can be accelerated by general techniques designed for kernel-based learning. We provide explicit analytical expressions for our kernel and argue that our formulation can be seen as a generalization of cubic spline interpolation to higher dimensions. In particular, the RKHS norm associated with Neural Splines biases toward smooth interpolants.

Super-resolution based on deep convolutional networks is currently gaining much attention from both academia and industry. However, lack of proper evaluation measures makes it difficult to compare approaches, hampering progress in the field. Traditional measures, such as PSNR or SSIM, are known to poorly correlate with the human perception of image quality. Therefore, in existing works common practice is also to report Mean-Opinion-Score (MOS) -- the results of human evaluation of super-resolved images. Unfortunately, the MOS values from different papers are not directly comparable, due to the varying number of raters, their subjectivity, etc. By this paper, we introduce Neural Side-By-Side -- a new measure that allows super-resolution models to be compared automatically, effectively approximating human preferences. Namely, we collect a large dataset of aligned image pairs, which were produced by different super-resolution models. Then each pair is annotated by several raters, who were instructed to choose a more visually appealing image. Given the dataset and the labels, we trained a CNN model that obtains a pair of images and for each image predicts a probability of being more preferable than its counterpart. In this work, we show that Neural Side-By-Side generalizes across both new models and new data. Hence, it can serve as a natural approximation of human preferences, which can be used to compare models or tune hyperparameters without raters' assistance. We open-source the dataset and the pretrained model and expect that it will become a handy tool for researchers and practitioners.

Recent implicit neural rendering methods have demonstrated that it is possible to learn accurate view synthesis for complex scenes by predicting their volumetric density and color supervised solely by a set of RGB images. However, existing methods are restricted to learning efficient representations of static scenes that encode all scene objects into a single neural network, and they lack the ability to represent dynamic scenes and decompose scenes into individual objects. In this work, we present the first neural rendering method that represents multi-object dynamic scenes as scene graphs. We propose a learned scene graph representation, which encodes object transformations and radiance, allowing us to efficiently render novel arrangements and views of the scene. To this end, we learn implicitly encoded scenes, combined with a jointly learned latent representation to describe similar objects with a single implicit function. We assess the proposed method on synthetic and real automotive data, validating that our approach learns dynamic scenes -- only by observing a video of this scene -- and allows for rendering novel photo-realistic views of novel scene compositions with unseen sets of objects at unseen poses.

We present a method to perform novel view and time synthesis of dynamic scenes, requiring only a monocular video with known camera poses as input. To do this, we introduce Neural Scene Flow Fields, a new representation that models the dynamic scene as a time-variant continuous function of appearance, geometry, and 3D scene motion. Our representation is optimized through a neural network to fit the observed input views. We show that our representation can be used for complex dynamic scenes, including thin structures, view-dependent effects, and natural degrees of motion. We conduct a number of experiments that demonstrate our approach significantly outperforms recent monocular view synthesis methods, and show qualitative results of space-time view synthesis on a variety of real-world videos.

Is critical input information encoded in specific sparse pathways within the neural network? In this work, we discuss the problem of identifying these critical pathways and subsequently leverage them for interpreting the network's response to an input. The pruning objective --- selecting the smallest group of neurons for which the response remains equivalent to the original network --- has been previously proposed for identifying critical pathways. We demonstrate that sparse pathways derived from pruning do not necessarily encode critical input information. To ensure sparse pathways include critical fragments of the encoded input information, we propose pathway selection via neurons' contribution to the response. We proceed to explain how critical pathways can reveal critical input features. We prove that pathways selected via neuron contribution are locally linear (in an L2-ball), a property that we use for proposing a feature attribution method: "pathway gradient". We validate our interpretation method using mainstream evaluation experiments. The validation of pathway gradient interpretation method further confirms that selected pathways using neuron contributions correspond to critical input features. The code is publicly available.

Absolute camera pose estimation is usually addressed by sequentially solving two distinct subproblems: First a feature matching problem that seeks to establish putative 2D-3D correspondences, and then a Perspective-n-Point problem that minimizes, w.r.t. the camera pose, the sum of so-called Reprojection Errors (RE). We argue that generating putative 2D-3D correspondences 1) leads to an important loss of information that needs to be compensated as far as possible, within RE, through the choice of a robust loss and the tuning of its hyperparameters and 2) may lead to an RE that conveys erroneous data to the pose estimator. In this paper, we introduce the Neural Reprojection Error (NRE) as a substitute for RE. NRE allows to rethink the camera pose estimation problem by merging it with the feature learning problem, hence leveraging richer information than 2D-3D correspondences and eliminating the need for choosing a robust loss and its hyperparameters. Thus NRE can be used as training loss to learn image descriptors tailored for pose estimation. We also propose a coarse-to-fine optimization method able to very efficiently minimize a sum of NRE terms w.r.t. the camera pose. We experimentally demonstrate that NRE is a good substitute for RE as it significantly improves both the robustness and the accuracy of the camera pose estimate while being computationally and memory highly efficient. From a broader point of view, we believe this new way of merging deep learning and 3D geometry may be useful in other computer vision applications.

Prototype-based methods use interpretable representations to address the black-box nature of deep learning models, in contrast to post-hoc explanation methods that only approximate such models. We propose the Neural Prototype Tree (ProtoTree), an intrinsically interpretable deep learning method for fine-grained image recognition. ProtoTree combines prototype learning with decision trees, and thus results in a globally interpretable model by design. Additionally, ProtoTree can locally explain a single prediction by outlining a decision path through the tree. Each node in our binary tree contains a trainable prototypical part. The presence or absence of this learned prototype in an image determines the routing through a node. Decision making is therefore similar to human reasoning: Does the bird have a red throat? And an elongated beak? Then it's a hummingbird! We tune the accuracy-interpretability trade-off using ensemble methods, pruning and binarizing. We apply pruning without sacrificing accuracy, resulting in a small tree with only 8 learned prototypes along a path to classify a bird from 200 species. An ensemble of 5 ProtoTrees achieves competitive accuracy on the CUB-200- 2011 and Stanford Cars data sets. Code is available at https://github.com/M-Nauta/ProtoTree.

Impressive progress in 3D shape extraction led to representations that can capture object geometries with high fidelity. In parallel, primitive-based methods seek to represent objects as semantically consistent part arrangements. However, due to the simplicity of existing primitive representations, these methods fail to accurately reconstruct 3D shapes using a small number of primitives/parts. We address the trade-off between reconstruction quality and number of parts with Neural Parts, a novel 3D primitive representation that defines primitives using an Invertible Neural Network (INN) which implements homeomorphic mappings between a sphere and the target object. The INN allows us to compute the inverse mapping of the homeomorphism, which in turn, enables the efficient computation of both the implicit surface function of a primitive and its mesh, without any additional post-processing. Our model learns to parse 3D objects into semantically consistent part arrangements without any part-level supervision.Evaluations on ShapeNet, D-FAUST and FreiHAND demonstrate that our primitives can capture complex geometries and thus simultaneously achieve geometrically accurate as well as interpretable reconstructions using an order of magnitude fewer primitives than state-of-the-art shape abstraction methods.