CVPR 2021 Papers

We present a method for reconstructing triangle meshes from point clouds. Existing learning-based methods for mesh reconstruction mostly generate triangles individually, making it hard to create manifold meshes. We leverage the properties of 2D Delaunay triangulations to construct a mesh from manifold surface elements. Our method first estimates local geodesic neighborhoods around each point. We then perform a 2D projection of these neighborhoods using a learned logarithmic map. A Delaunay triangulation in this 2D domain is guaranteed to produce a manifold patch, which we call a surface element. We synchronize the local 2D projections of neighboring elements to maximize the manifoldness of the reconstructed mesh. Our results show that we achieve better overall manifoldness of our reconstructed meshes than current methods to reconstruct meshes with arbitrary topology. Our code, data and pretrained models can be found online: https://github.com/mrakotosaon/dse-meshing

Learning latent variable models with deep top-down architectures typically requires inferring the latent variables for each training example based on the posterior distribution of these latent variables. The inference step typically relies on either time-consuming long run Markov chain Monte Caro (MCMC) or a separate inference model for variational learning. In this paper, we propose to use short run MCMC, such as Langevin dynamics, as an approximate inference engine, where the bias existing in the output distribution of the short run Langevin dynamics is corrected by optimal transport, which aims at minimizing the Wasserstein distance between the biased distribution produced by the finite step Langevin dynamics and the prior distribution. Our experiments show that the proposed strategy outperforms the variational auto-encoder (VAE) and alternating back-propagation algorithm (ABP) in terms of reconstruction error and synthesis quality.

The problem of novelty detection in fine-grained visual classification (FGVC) is considered. An integrated understanding of the probabilistic and distance-based approaches to novelty detection is developed within the framework of convolutional neural networks (CNNs). It is shown that softmax CNN classifiers are inconsistent with novelty detection, because their learned class-conditional distributions and associated distance metrics are unidentifiable. A new regularization constraint, the class-conditional Gaussianity loss, is then proposed to eliminate this unidentifiability, and enforce Gaussian class-conditional distributions. This enables training Novelty Detection Consistent Classifiers (NDCCs) that are jointly optimal for classification and novelty detection. Empirical evaluations show that NDCCs achieve significant improvements over the state-of-the-art on both small- and large-scale FGVC datasets.

Integrated interpretability without sacrificing the prediction accuracy of decision making algorithms has the potential of greatly improving their value to the user. Instead of assigning a label to an image directly, we propose to learn iterative binary sub-decisions, inducing sparsity and transparency in the decision making process. The key aspect of our model is its ability to build a decision tree whose structure is encoded into the memory representation of a Recurrent Neural Network jointly learned by two models communicating through message passing. In addition, our model assigns a semantic meaning to each decision in the form of binary attributes, providing concise, semantic and relevant rationalizations to the user. On three benchmark image classification datasets, including the large-scale ImageNet, our model generates human interpretable binary decision sequences explaining the predictions of the network while maintaining state-of-the-art accuracy.

Recently, cross-modal retrieval is emerging with the help of deep multimodal learning. However, even for unimodal data, collecting large-scale well-annotated data is expensive and time-consuming, and not to mention the additional challenges from multiple modalities. Although crowd-sourcing annotation, e.g., Amazon's Mechanical Turk, can be utilized to mitigate the labeling cost, but leading to the unavoidable noise in labels for the non-expert annotating. To tackle the challenge, this paper presents a general Multimodal Robust Learning framework (MRL) for learning with multimodal noisy labels to mitigate noisy samples and correlate distinct modalities simultaneously. To be specific, we propose a Robust Clustering loss (RC) to make the deep networks focus on clean samples instead of noisy ones. Besides, a simple yet effective multimodal loss function, called Multimodal Contrastive loss (MC), is proposed to maximize the mutual information between different modalities, thus alleviating the interference of noisy samples and cross-modal discrepancy. Extensive experiments are conducted on four widely-used multimodal datasets to demonstrate the effectiveness of the proposed approach by comparing to 14 state-of-the-art methods.

How to represent an image? While the visual world is presented in a continuous manner, machines store and see the images in a discrete way with 2D arrays of pixels. In this paper, we seek to learn a continuous representation for images. Inspired by the recent progress in 3D reconstruction with implicit neural representation, we propose Local Implicit Image Function (LIIF), which takes an image coordinate and the 2D deep features around the coordinate as inputs, predicts the RGB value at a given coordinate as an output. Since the coordinates are continuous, LIIF can be presented in arbitrary resolution. To generate the continuous representation for images, we train an encoder with LIIF representation via a self-supervised task with super-resolution. The learned continuous representation can be presented in arbitrary resolution even extrapolate to x30 higher resolution, where the training tasks are not provided. We further show that LIIF representation builds a bridge between discrete and continuous representation in 2D, it naturally supports the learning tasks with size-varied image ground-truths and significantly outperforms the method with resizing the ground-truths.

Learning based representation has become the key to the success of many computer vision systems. While many 3D representations have been proposed, it is still an unaddressed problem how to represent a dynamically changing 3D object. In this paper, we introduce a compositional representation for 4D captures, i.e. a deforming 3D object over a temporal span, that disentangles shape, initial state, and motion respectively. Each component is represented by a latent code via a trained encoder. To model the motion, a neural Ordinary Differential Equation (ODE) is trained to update the initial state conditioned on the learned motion code, and a decoder takes the shape code and the updated state code to reconstruct the 3D model at each time stamp. To this end, we propose an Identity Exchange Training (IET) strategy to encourage the network to learn effectively decoupling each component. Extensive experiments demonstrate that the proposed method outperforms existing state-of-the-art deep learning based methods on 4D reconstruction, and significantly improves on various tasks, including motion transfer and completion.

Photorealistic rendering of dynamic humans is an important ability for telepresence systems, virtual shopping, synthetic data generation, and more. Recently, neural rendering methods, which combine techniques from computer graphics and machine learning, have created high-fidelity models of humans and objects. Some of these methods do not produce results with high-enough fidelity for driveable human models (Neural Volumes) whereas others have extremely long rendering times (NeRF). We propose a novel compositional 3D representation that combines the best of previous methods to produce both higher-resolution and faster results. Our representation bridges the gap between discrete and continuous volumetric representations by combining a coarse 3D-structure-aware grid of animation codes with a continuous learned scene function that maps every position and its corresponding local animation code to its view-dependent emitted radiance and local volume density. Differentiable volume rendering is employed to compute photo-realistic novel views of the human head and upper body as well as to train our novel representation end-to-end using only 2D supervision. In addition, we show that the learned dynamic radiance field can be used to synthesize novel unseen expressions based on a global animation code. Our approach achieves state-of-the-art results for synthesizing novel views of dynamic human heads and the upper body. See our project page for more results.

Most 3D face reconstruction methods rely on 3D morphable models, which disentangle the space of facial deformations into identity and expression geometry, and skin reflectance. These models are typically learned from a limited number of 3D scans and thus do not generalize well across different identities and expressions. We present the first approach to learn complete 3D models of face identity and expression geometry, and reflectance, just from images and videos. The virtually endless collection of such data, in combination with our self-supervised learning-based approach allows for learning face models that generalize beyond the span of existing approaches. Our network design and loss functions ensure a disentangled parameterization of not only identity and albedo, but also, for the first time, an expression basis. Our method also allows for in-the-wild monocular reconstruction at test time. We show that our learned models better generalize and lead to higher quality image-based reconstructions than existing approaches. We show that the learned model can also be personalized to a video, for a better capture of the geometry and albedo.

Camera localization aims to estimate 6 DoF camera poses from RGB images. Traditional methods detect and match interest points between a query image and a pre-built 3D model. Recent learning-based approaches encode scene structures into a specific convolutional neural network(CNN) and thus are able to predict dense coordinates from RGB images. However, most of them require re-training or re-adaption for a new scene and have difficulties in handling large-scale scenes due to limited network capacity. We present a new method for scene agnostic camera localization using dense scene matching (DSM), where the cost volume is constructed between a query image and a scene. The cost volume and the corresponding coordinates are processed by a CNN to predict dense coordinates. Camera poses can then be solved by PnP algorithms. In addition, our method can be extended to temporal domain, giving extra performance boost during testing time. Our scene-agnostic approach achieves comparable accuracy as the existing scene-specific approaches on the 7scenes and Cambridge benchmark. This approach also remarkably outperforms state-of-the-art scene-agnostic dense coordinate regression network SANet.

In medical image analysis, it is typical to collect multiple annotations, each from a different clinical expert or rater, in the expectation that possible diagnostic errors could be mitigated. Meanwhile, from the computer vision practitioner viewpoint, it has been a common practice to adopt the ground-truth obtained via either the majority-vote or simply one annotation from a preferred rater. This process, however, tends to overlook the rich information of agreement or disagreement ingrained in the raw multi-rater annotations. To address this issue, we propose to explicitly model the multi-rater (dis-)agreement, dubbed MRNet, which has two main contributions. First, an expertise-aware inferring module or EIM is devised to embed the expertise level of individual raters as prior knowledge, to form high-level semantic features. Second, our approach is capable of reconstructing multi-rater gradings from coarse predictions, with the multi-rater (dis-)agreement cues being further exploited to improve the segmentation performance. To our knowledge, our work is the first in producing calibrated predictions under different expertise levels for medical image segmentation. Extensive empirical experiments are conducted across five medical segmentation tasks of diverse imaging modalities. In these experiments, superior performance of our MRNet is observed comparing to the state-of-the-arts, indicating the effectiveness and applicability of our MRNet toward a wide range of medical segmentation tasks.

GuessWhat?! is a visual dialog guessing game which incorporates a Questioner agent that generates a sequence of questions, while an Oracle agent answers the respective questions about a target object in an image. Based on this dialog history between the Questioner and the Oracle, a Guesser agent makes a final guess of the target object. While previous work has focused on dialogue policy optimization and visual-linguistic information fusion, most work learns the vision-linguistic encoding for the three agents solely on the GuessWhat?! dataset without shared and prior knowledge of vision-linguistic representation. To bridge these gaps, this paper proposes new Oracle, Guesser and Questioner models that take advantage of a pretrained vision-linguistic model, VilBert. For Oracle model, we introduce a two-way background/target fusion mechanism to understand both intra and inter-object questions. For Guesser model, we introduce a state-estimator that best utilizes Vilbert's strength in single-turn referring expression comprehension. For the Questioner, we share the state-estimator from pretrained Guesser with Questioner to guide the question generator. Experimental results show that our proposed models outperform state-of-the-art models significantly by 7%, 10%, 12% for Oracle, Guesser and End-to-End Questioner respectively.

Human activities can be learned from video. With effective modeling it is possible to discover not only the action labels but also the temporal structure of the activities, such as the progression of the sub-activities. Automatically recognizing such structure from raw video signal is a new capability that promises authentic modeling and successful recognition of human-object interactions. Toward this goal, we introduce Asynchronous-Sparse Interaction Graph Networks (ASSIGN), a recurrent graph network that is able to automatically detect the structure of interaction events associated with entities in a video scene. ASSIGN pioneers learning of autonomous behavior of video entities including their dynamic structure and their interaction with the coexisting neighbors. Entities' lives in our model are asynchronous to those of others therefore more flexible in adapting to complex scenarios. Their interactions are sparse in time hence more faithful to the true underlying nature and more robust in inference and learning. ASSIGN is tested on human-object interaction recognition and shows superior performance in segmenting and labeling of human sub-activities and object affordances from raw videos. The native ability of ASSIGN in discovering temporal structure also eliminates the dependence on external segmentation that was previously mandatory for this task.

We show that learning affinity in upsampling provides an effective and efficient approach to exploit pairwise interactions in deep networks. Second-order features are commonly used in dense prediction to build adjacent relations with a learnable module after upsampling such as non-local blocks. Since upsampling is essential, learning affinity in upsampling can avoid additional propagation layers, offering the potential for building compact models. By looking at existing upsampling operators from a unified mathematical perspective, we generalize them into a second-order form and introduce Affinity-Aware Upsampling (A2U) where upsampling kernels are generated using a light-weight low-rank bilinear model and are conditioned on second-order features. Our upsampling operator can also be extended to downsampling. We discuss alternative implementations of A2U and verify their effectiveness on two detail-sensitive tasks: image reconstruction on a toy dataset; and a large-scale image matting task where affinity-based ideas constitute mainstream matting approaches. In particular, results on the Composition-1k matting dataset show that A2U achieves a 14% relative improvement in the SAD metric against a strong baseline with negligible increase of parameters (< 0.5%). Compared with the state-of-the-art matting network, we achieve 8% higher performance with only 40% model complexity.

Establishing dense correspondences between a pair of images is an important and general problem. However, dense flow estimation is often inaccurate in the case of large displacements or homogeneous regions. For most applications and down-stream tasks, such as pose estimation, image manipulation, or 3D reconstruction, it is crucial to know when and where to trust the estimated matches. In this work, we aim to estimate a dense flow field relating two images, coupled with a robust pixel-wise confidence map indicating the reliability and accuracy of the prediction. We develop a flexible probabilistic approach that jointly learns the flow prediction and its uncertainty. In particular, we parametrize the predictive distribution as a constrained mixture model, ensuring better modelling of both accurate flow predictions and outliers. Moreover, we develop an architecture and training strategy tailored for robust and generalizable uncertainty prediction in the context of self-supervised training. Our approach obtains state-of-the-art results on multiple challenging geometric matching and optical flow datasets. We further validate the usefulness of our probabilistic confidence estimation for the task of pose estimation. Code and models are available at https://github.com/PruneTruong/PDCNet.

It is well acknowledged that person re-identification (person ReID) highly relies on visual texture information like clothing. Despite significant progress has been made in recent years, texture-confusing situations like clothing changing and persons wearing the same clothes receive little attention from most existing ReID methods. In this paper, rather than relying on texture based information, we propose to improve the robustness of person ReID against clothing texture by exploiting the information of a person's 3D shape. Existing shape learning schemas for person ReID either ignore the 3D information of a person, or require extra physical devices to collect 3D source data. Differently, we propose a novel ReID learning framework that directly extracts a texture-insensitive 3D shape embedding from a 2D image by adding 3D body reconstruction as an auxiliary task and regularization, called 3D Shape Learning (3DSL). The 3D reconstruction based regularization forces the ReID model to decouple the 3D shape information from the visual texture, and acquire discriminative 3D shape ReID features. To solve the problem of lacking 3D ground truth, we design an adversarial self-supervised projection (ASSP) model, performing 3D reconstruction without ground truth. Extensive experiments on common ReID datasets and texture-confusing datasets validate the effectiveness of our model.

Coordinate-based neural representations have shown significant promise as an alternative to discrete, array-based representations for complex low dimensional signals. However, optimizing a coordinate-based network from randomly initialized weights for each new signal is inefficient. We propose applying standard meta-learning algorithms to learn the initial weight parameters for these fully-connected networks based on the underlying class of signals being represented (e.g., images of faces or 3D models of chairs). Despite requiring only a minor change in implementation, using these learned initial weights enables faster convergence during optimization and can serve as a strong prior over the signal class being modeled, resulting in better generalization when only partial observations of a given signal are available. We explore these benefits across a variety of tasks, including representing 2D images, reconstructing CT scans, and recovering 3D shapes and scenes from 2D image observations.

Motion coherence is an important clue for distinguishing true correspondences from false ones. Modeling motion coherence on sparse putative correspondences is challenging due to their sparsity and uneven distributions. Existing works on motion coherence are sensitive to parameter settings and have difficulty in dealing with complex motion patterns. In this paper, we introduce a network called Laplacian Motion Coherence Network (LMCNet) to learn motion coherence property for correspondence pruning. We propose a novel formulation of fitting coherent motions with a smooth function on a graph of correspondences and show that this formulation allows a closed-form solution by graph Laplacian. This closed-form solution enables us to design a differentiable layer in a learning framework to capture global motion coherence from putative correspondences. The global motion coherence is further combined with local coherence extracted by another local layer to robustly detect inlier correspondences. Experiments demonstrate that LMCNet has superior performances to the state of the art in relative camera pose estimation and correspondences pruning of dynamic scenes.

Data association across frames is at the core of Multiple Object Tracking (MOT) task. This problem is usually solved by a traditional graph-based optimization or directly learned via deep learning. Despite their popularity, we find some points worth studying in current paradigm: 1) Existing methods mostly ignore the context information among tracklets and intra-frame detections, which makes the tracker hard to survive in challenging cases like severe occlusion. 2) The end-to-end association methods solely rely on the data fitting power of deep neural networks, while they hardly utilize the advantage of optimization-based assignment methods. 3) The graph-based optimization methods mostly utilize a separate neural network to extract features, which brings the inconsistency between training and inference. Therefore, in this paper we propose a novel learnable graph matching method to address these issues. Briefly speaking, we model the relationships between tracklets and the intra-frame detections as a general undirected graph. Then the association problem turns into a general graph matching between tracklet graph and detection graph. Furthermore, to make the optimization end-to-end differentiable, we relax the original graph matching into continuous quadratic programming and then incorporate the training of it into a deep graph network with the help of the implicit function theorem. Lastly, our method GMTracker, achieves state-of-the-art performance on several standard MOT datasets. Our code is available at https://github.com/jiaweihe1996/GMTracker.