CVPR 2020 Papers

Current 6D object pose estimation methods usually require a 3D model for each object. These methods also require additional training in order to incorporate new objects. As a result, they are difficult to scale to a large number of objects and cannot be directly applied to unseen objects. We propose a novel framework for 6D pose estimation of unseen objects. We present a network that reconstructs a latent 3D representation of an object using a small number of reference views at inference time. Our network is able to render the latent 3D representation from arbitrary views. Using this neural renderer, we directly optimize for pose given an input image. By training our network with a large number of 3D shapes for reconstruction and rendering, our network generalizes well to unseen objects. We present a new dataset for unseen object pose estimation--MOPED. We evaluate the performance of our method for unseen object pose estimation on MOPED as well as the ModelNet and LINEMOD datasets. Our method performs competitively to supervised methods that are trained on those objects. Code and data will be available at https://keunhong.com/publications/latentfusion/

Representation learning is widely applied for various tasks on multimedia data, e.g., retrieval and search. One approach for learning useful representation is by utilizing the relationships or similarities between examples. In this work, we explore two promising scalable representation learning approaches on video domain. With hierarchical graph clusters built upon video-to-video similarities, we propose: 1) smart negative sampling strategy that significantly boosts training efficiency with triplet loss, and 2) a pseudo-classification approach using the clusters as pseudo-labels. The embeddings trained with the proposed methods are competitive on multiple video understanding tasks, including related video retrieval and video annotation. Both of these proposed methods are highly scalable, as verified by experiments on large-scale datasets.

Many existing studies reveal that annotation inconsistency widely exists among a variety of facial expression recognition (FER) datasets. The reason might be the subjectivity of human annotators and the ambiguous nature of the expression labels. One promising strategy tackling such a problem is a recently proposed learning paradigm called Label Distribution Learning (LDL), which allows multiple labels with different intensity to be linked to one expression. However, it is often impractical to directly apply label distribution learning because numerous existing datasets only contain one-hot labels rather than label distributions. To solve the problem, we propose a novel approach named Label Distribution Learning on Auxiliary Label Space Graphs(LDL-ALSG) that leverages the topological information of the labels from related but more distinct tasks, such as action unit recognition and facial landmark detection. The underlying assumption is that facial images should have similar expression distributions to their neighbours in the label space of action unit recognition and facial landmark detection. Our proposed method is evaluated on a variety of datasets and outperforms those state-of-the-art methods consistently with a huge margin.

To get more accurate saliency maps, recent methods mainly focus on aggregating multi-level features from fully convolutional network (FCN) and introducing edge information as auxiliary supervision. Though remarkable progress has been achieved, we observe that the closer the pixel is to the edge, the more difficult it is to be predicted, because edge pixels have a very imbalance distribution. To address this problem, we propose a label decoupling framework (LDF) which consists of a label decoupling (LD) procedure and a feature interaction network (FIN). LD explicitly decomposes the original saliency map into body map and detail map, where body map concentrates on center areas of objects and detail map focuses on regions around edges. Detail map works better because it involves much more pixels than traditional edge supervision. Different from saliency map, body map discards edge pixels and only pays attention to center areas. This successfully avoids the distraction from edge pixels during training. Therefore, we employ two branches in FIN to deal with body map and detail map respectively. Feature interaction (FI) is designed to fuse the two complementary branches to predict the saliency map, which is then used to refine the two branches again. This iterative refinement is helpful for learning better representations and more precise saliency maps. Comprehensive experiments on six benchmark datasets demonstrate that LDF outperforms state-of-the-art approaches on different evaluation metrics.

Modern face alignment methods have become quite accurate at predicting the locations of facial landmarks, but they do not typically estimate the uncertainty of their predicted locations nor predict whether landmarks are visible. In this paper, we present a novel framework for jointly predicting landmark locations, associated uncertainties of these predicted locations, and landmark visibilities. We model these as mixed random variables and estimate them using a deep network trained using our proposed Location, Uncertainty, and Visibility Likelihood (LUVLi) loss. In addition, we release an entirely new labeling of a large face alignment dataset with over 19,000 face images in a full range of head poses. Each face is manually labeled with the ground-truth locations of 68 landmarks, with the additional information of whether each landmarks is visible, self-occluded (due to extreme head poses), or externally occluded. Not only does our joint estimation yield accurate estimates of the uncertainty of predicted landmark locations, but it also yields state-of-the-art estimates for the landmark locations themselves on mulitple standard face alignment datasets. Our method's estimates of the uncertainty of predicted landmark locations could be used to automatically identify input images on which face alignment fails, which can be critical for downstream tasks.

We introduce a one-shot segmentation method to alleviate the burden of manual annotation for medical images. The main idea is to treat one-shot segmentation as a classical atlas-based segmentation problem, where voxel-wise correspondence from the atlas to the unlabelled data is learned. Subsequently, segmentation label of the atlas can be transferred to the unlabelled data with the learned correspondence. However, since ground truth correspondence between images is usually unavailable, the learning system must be well-supervised to avoid mode collapse and convergence failure. To overcome this difficulty, we resort to the forward-backward consistency, which is widely used in correspondence problems, and additionally learn the backward correspondences from the warped atlases back to the original atlas. This cycle-correspondence learning design enables a variety of extra, cycle-consistency-based supervision signals to make the training process stable, while also boost the performance. We demonstrate the superiority of our method over both deep learning-based one-shot segmentation methods and a classical multi-atlas segmentation method via thorough experiments.

Deep neural networks have made tremendous progress in 3D point-cloud recognition. Recent works have shown that these 3D recognition networks are also vulnerable to adversarial samples produced from various attack methods, including optimization-based 3D Carlini-Wagner attack, gradient-based iterative fast gradient method, and skeleton-detach based point-dropping. However, after a careful analysis, these methods are either extremely slow because of the optimization/iterative scheme, or not flexible to support targeted attack of a specific category. To overcome these shortcomings, this paper proposes a novel label guided adversarial network (LG-GAN) for real-time flexible targeted point cloud attack. To the best of our knowledge, this is the first generation based 3D point cloud attack method. By feeding the original point clouds and target attack label into LG-GAN, it can learn how to deform the point clouds to mislead the recognition network into the specific label only with a single forward pass. In detail, LG-GAN first leverages one multi-branch adversarial network to extract hierarchical features of the input point clouds, then incorporates the specified label information into multiple intermediate features using the label encoder. Finally, the encoded features will be fed into the coordinate reconstruction decoder to generate the target adversarial sample. By evaluating different point-cloud recognition models (e.g., PointNet, PointNet++ and DGCNN), we demonstrate that the proposed LG-GAN can support flexible targeted attack on the fly while guaranteeing good attack performance and higher efficiency simultaneously.

Graph convolution networks (GCN) are increasingly popular in many applications, yet remain notoriously hard to train over large graph datasets. They need to compute node representations recursively from their neighbors. Current GCN training algorithms suffer from either high computational costs that grow exponentially with the number of layers, or high memory usage for loading the entire graph and node embeddings. In this paper, we propose a novel efficient layer-wise training framework for GCN (L-GCN), that disentangles feature aggregation and feature transformation during training, hence greatly reducing time and memory complexities. We present theoretical analysis for L-GCN under the graph isomorphism framework, that L-GCN leads to as powerful GCNs as the more costly conventional training algorithm does, under mild conditions. We further propose L^2-GCN, which learns a controller for each layer that can automatically adjust the training epochs per layer in L-GCN. Experiments show that L-GCN is faster than state-of-the-arts by at least an order of magnitude, with a consistent of memory usage not dependent on dataset size, while maintaining comparable prediction performance. With the learned controller, L^2-GCN can further cut the training time in half. Our codes are available at https://github.com/Shen-Lab/L2-GCN.

Cross-modal knowledge distillation deals with transferring knowledge from a model trained with superior modalities (Teacher) to another model trained with weak modalities (Student). Existing approaches require paired training examples exist in both modalities. However, accessing the data from superior modalities may not always be feasible. For example, in the case of 3D hand pose estimation, depth maps, point clouds, or stereo images usually capture better hand structures than RGB images, but most of them are expensive to be collected. In this paper, we propose a novel scheme to train the Student in a Target dataset where the Teacher is unavailable. Our key idea is to generalize the distilled cross-modal knowledge learned from a Source dataset, which contains paired examples from both modalities, to the Target dataset by modeling knowledge as priors on parameters of the Student. We name our method "Cross-Modal Knowledge Generalization" and demonstrate that our scheme results in competitive performance for 3D hand pose estimation on standard benchmark datasets.

Detecting 3D objects keypoints is ofgreat interest to the areas of both graphics and computer vision. There have been several 2D and 3D keypoint datasets aiming to address this problem in a data-driven way. These datasets, however, either lack scalability or bring ambiguity to the definition of keypoints. Therefore, we present KeypointNet: the first large-scale and diverse 3D keypoint dataset that contains 83,231 keypoints and 8,329 3D models from 16 object categories, by leveraging numerous human annotations. To handle the inconsistency between annotations from different people, we propose a novel method to aggregate these keypoints automatically, through minimization of a fidelity loss. Finally, ten state-of-the-art methods are benchmarked on our proposed dataset.

Estimating the 3D pose of desktop objects is crucial for applications such as robotic manipulation. Many existing approaches to this problem require a depth map of the object for both training and prediction, which restricts them to opaque, lambertian objects that produce good returns in an RGBD sensor. In this paper we forgo using a depth sensor in favor of raw stereo input. We address two problems: first, we establish an easy method for capturing and labeling 3D keypoints on desktop objects with an RGB camera; and second, we develop a deep neural network, called KeyPose, that learns to accurately predict object poses using 3D keypoints, from stereo input, and works even for transparent objects. To evaluate the performance of our method, we create a dataset of 15 clear objects in five classes, with 48K 3D-keypoint labeled images. We train both instance and category models, and show generalization to new textures, poses, and objects. KeyPose surpasses state-of-the-art performance in 3D pose estimation on this dataset by factors of 1.5 to 3.5, even in cases where the competing method is provided with ground-truth depth. Stereo input is essential for this performance as it improves results compared to using monocular input by a factor of 2. We will release a public version of the data capture and labeling pipeline, the transparent object database, and the KeyPose models and evaluation code. Project website: https://sites.google.com/corp/view/keypose.

This paper proposes a joint training method to learn both the variational auto-encoder (VAE) and the latent energy-based model (EBM). The joint training of VAE and latent EBM are based on an objective function that consists of three Kullback-Leibler divergences between three joint distributions on the latent vector and the image, and the objective function is of an elegant symmetric and anti-symmetric form of divergence triangle that seamlessly integrates variational and adversarial learning. In this joint training scheme, the latent EBM serves as a critic of the generator model, while the generator model and the inference model in VAE serve as the approximate synthesis sampler and inference sampler of the latent EBM. Our experiments show that the joint training greatly improves the synthesis quality of the VAE. It also enables learning of an energy function that is capable of detecting out of sample examples for anomaly detection.

Due to inevitable noises and quantization error, the reconstructed 3D models via RGB-D sensors always accompany geometric error and camera drifting, which consequently lead to blurring and unnatural texture mapping results. Most of the 3D reconstruction methods focus on either geometry refinement or texture improvement respectively, which subjectively decouples the inter-relationship between geometry and texture. In this paper, we propose a novel approach that can jointly optimize the camera poses, texture and geometry of the reconstructed model, and color consistency between the key-frames. Instead of computing Shape-From-Shading (SFS) expensively, our method directly optimizes the reconstructed mesh according to color and geometric consistency and high-boost normal cues, which can effectively overcome the texture-copy problem generated by SFS and achieve more detailed shape reconstruction. As the joint optimization involves multiple correlated terms, therefore, we further introduce an iterative framework to interleave the optimal state. The experiments demonstrate that our method can recover not only fine-scale geometry but also high-fidelity texture.

Directly learning multiple 3D objects motion from sequential images is difficult, while the geometric bundle adjustment lacks the ability to localize the invisible object centroid. To benefit from both the powerful object understanding skill from deep neural network meanwhile tackle precise geometry modeling for consistent trajectory estimation, we propose a joint spatial-temporal optimization-based stereo 3D object tracking method. From the network, we detect corresponding 2D bounding boxes on adjacent images and regress an initial 3D bounding box. Dense object cues (local depth and local coordinates) that associating to the object centroid are then predicted using a region-based network. Considering both the instant localization accuracy and motion consistency, our optimization models the relations between the object centroid and observed cues into a joint spatial-temporal error function. All historic cues will be summarized to contribute to the current estimation by a per-frame marginalization strategy without repeated computation. Quantitative evaluation on the KITTI tracking dataset shows our approach outperforms previous image-based 3D tracking methods by significant margins. We also report extensive results on multiple categories and larger datasets (KITTI raw and Argoverse Tracking) for future benchmarking.

In this paper, we present a joint multi-task learning framework for semantic segmentation and boundary detection. The critical component in the framework is the iterative pyramid context module (PCM), which couples two tasks and stores the shared latent semantics to interact between the two tasks. For semantic boundary detection, we propose the novel spatial gradient fusion to suppress non-semantic edges. As semantic boundary detection is the dual task of semantic segmentation, we introduce a loss function with boundary consistency constraint to improve the boundary pixel accuracy for semantic segmentation. Our extensive experiments demonstrate superior performance over state-of-the-art works, not only in semantic segmentation but also in semantic boundary detection. In particular, a mean IoU score of 81.8% on Cityscapes test set is achieved without using coarse data or any external data for semantic segmentation. For semantic boundary detection, we improve over previous state-of-the-art works by 9.9% in terms of AP and 6.8% in terms of MF(ODS).

Depth estimation is an essential component in understanding the 3D geometry of a scene, with numerous applications in urban and indoor settings. These scenarios are characterized by a prevalence of human made structures, which in most of the cases are either inherently piece-wise planar or can be approximated as such. With these settings in mind, we devise a novel depth refinement framework that aims at recovering the underlying piece-wise planarity of those inverse depth maps associated to piece-wise planar scenes. We formulate this task as an optimization problem involving a data fidelity term, which minimizes the distance to the noisy and possibly incomplete input inverse depth map, as well as a regularization, which enforces a piece-wise planar solution. As for the regularization term, we model the inverse depth map pixels as the nodes of a weighted graph, with the weight of the edge between two pixels capturing the likelihood that they belong to the same plane in the scene. The proposed regularization fits a plane at each pixel automatically, avoiding any a priori estimation of the scene planes, and enforces that strongly connected pixels are assigned to the same plane. The resulting optimization problem is solved efficiently with the ADAM solver. Extensive tests show that our method leads to a significant improvement in depth refinement, both visually and numerically, with respect to state-of-the-art algorithms on the Middlebury, KITTI and ETH3D multi-view datasets.

We present a novel computational imaging system with high resolution and low noise. Our system consists of a traditional video camera which captures high-resolution intensity images, and an event camera which encodes high-speed motion as a stream of asynchronous binary events. To process the hybrid input, we propose a unifying framework that first bridges the two sensing modalities via a noise-robust motion compensation model, and then performs joint image filtering. The filtered output represents the temporal gradient of the captured space-time volume, which can be viewed as motion-compensated event frames with high resolution and low noise. Therefore, the output can be widely applied to many existing event-based algorithms that are highly dependent on spatial resolution and noise robustness. In experimental results performed on both publicly available datasets as well as our contributing RGB-DAVIS dataset, we show systematic performance improvement in applications such as high frame-rate video synthesis, feature/corner detection and tracking, as well as high dynamic range image reconstruction.

Usually located at the very early stages of the computational photography pipeline, demosaicing and denoising play important parts in the modern camera image processing. Recently, some neural networks have shown the effectiveness in joint demosaicing and denoising (JDD). Most of them first decompose a Bayer raw image into a four-channel RGGB image and then feed it into a neural network. This practice ignores the fact that the green channels are sampled at a double rate compared to the red and the blue channels. In this paper, we propose a self-guidance network (SGNet), where the green channels are initially estimated and then works as a guidance to recover all missing values in the input image. In addition, as regions of different frequencies suffer different levels of degradation in image restoration. We propose a density-map guidance to help the model deal with a wide range of frequencies. Our model outperforms state-of-the-art joint demosaicing and denoising methods on four public datasets, including two real and two synthetic data sets. Finally, we also verify that our method obtains best results in joint demosaicing , denoising and super-resolution.

Currently, in Autonomous Driving (AD), most of the 3D object detection frameworks (either anchor- or anchor-free-based) consider the detection as a Bounding Box (BBox) regression problem. However, this compact representation is not sufficient to explore all the information of the objects. To tackle this problem, we propose a simple but practical detection framework to jointly predict the 3D BBox and instance segmentation. For instance segmentation, we propose a Spatial Embeddings (SEs) strategy to assemble all foreground points into their corresponding object centers. Base on the SE results, the object proposals can be generated based on a simple clustering strategy. For each cluster, only one proposal is generated. Therefore, the Non-Maximum Suppression (NMS) process is no longer needed here. Finally, with our proposed instance-aware ROI pooling, the BBox is refined by a second-stage network. Experimental results on the public KITTI dataset show that the proposed SEs can significantly improve the instance segmentation results compared with other feature embedding-based method. Meanwhile, it also outperforms most of the 3D object detectors on the KITTI testing benchmark.