CVPR 2019 Papers

Traditional 3D Convolutional Neural Networks (CNNs) are computationally expensive, memory intensive, prone to overfit, and most importantly, there is a need to improve their feature learning capabilities. To address these issues, we propose Rectified Local Phase Volume (ReLPV) block, an efficient alternative to the standard 3D convolutional layer. The ReLPV block extracts the phase in a 3D local neighborhood (e.g., 3x3x3) of each position of the input map to obtain the feature maps. The phase is extracted by computing 3D Short Term Fourier Transform (STFT) at multiple fixed low frequency points in the 3D local neighborhood of each position. These feature maps at different frequency points are then linearly combined after passing them through an activation function. The ReLPV block provides significant parameter savings of at least, 3^3 to 13^3 times compared to the standard 3D convolutional layer with the filter sizes 3x3x3 to 13x13x13, respectively. We show that the feature learning capabilities of the ReLPV block are significantly better than the standard 3D convolutional layer. Furthermore, it produces consistently better results across different 3D data representations. We achieve state-of-the-art accuracy on the volumetric ModelNet10 and ModelNet40 datasets while utilizing only 11% parameters of the current state-of-the-art. We also improve the state-of-the-art on the UCF-101 split-1 action recognition dataset by 5.68% (when trained from scratch) while using only 15% of the parameters of the state-of-the-art.

We present a novel deep convolutional network pipeline, LO-Net, for real-time lidar odometry estimation. Unlike most existing lidar odometry (LO) estimations that go through individually designed feature selection, feature matching, and pose estimation pipeline, LO-Net can be trained in an end-to-end manner. With a new mask-weighted geometric constraint loss, LO-Net can effectively learn feature representation for LO estimation, and can implicitly exploit the sequential dependencies and dynamics in the data. We also design a scan-to-map module, which uses the geometric and semantic information learned in LO-Net, to improve the estimation accuracy. Experiments on benchmark datasets demonstrate that LO-Net outperforms existing learning based approaches and has similar accuracy with the state-of-the-art geometry-based approach, LOAM.

We present LBS-AE; a self-supervised autoencoding algorithm for fitting articulated mesh models to point clouds. As input, we take a sequence of point clouds to be registered as well as an artist-rigged mesh, i.e. a template mesh equipped with a linear-blend skinning (LBS) deformation space parameterized by a skeleton hierarchy. As output, we learn an LBS-based autoencoder that produces registered meshes from the input point clouds. To bridge the gap between the artist-defined geometry and the captured point clouds, our autoencoder models pose-dependent deviations from the template geometry. During training, instead of us- ing explicit correspondences, such as key points or pose supervision, our method leverages LBS deformations to boot- strap the learning process. To avoid poor local minima from erroneous point-to-point correspondences, we utilize a structured Chamfer distance based on part-segmentations, which are learned concurrently using self-supervision. We demonstrate qualitative results on real captured hands, and report quantitative evaluations on the FAUST benchmark for body registration. Our method achieves performance that is superior to other unsupervised approaches and com- parable to methods using supervised examples.

Capturing the 'mutual gaze' of people is essential for understanding and interpreting the social interactions between them. To this end, this paper addresses the problem of detecting people Looking At Each Other (LAEO) in video sequences. For this purpose, we propose LAEO-Net, a new deep CNN for determining LAEO in videos. In contrast to previous works, LAEO-Net takes spatio-temporal tracks as input and reasons about the whole track. It consists of three branches, one for each character's tracked head and one for their relative position. Moreover, we introduce two new LAEO datasets: UCO-LAEO and AVA-LAEO. A thorough experimental evaluation demonstrates the ability of LAEO-Net to successfully determine if two people are LAEO and the temporal window where it happens. Our model achieves state-of-the-art results on the existing TVHID-LAEO video dataset, significantly outperforming previous approaches.

We present L3-Net - a novel learning-based LiDAR localization system that achieves centimeter-level localization accuracy, comparable to prior state-of-the-art systems with hand-crafted pipelines. Rather than relying on these hand-crafted modules, we innovatively implement the use of various deep neural network structures to establish a learning-based approach. L3-Net learns local descriptors specifically optimized for matching in different real-world driving scenarios. 3D convolutions over a cost volume built in the solution space significantly boosts the localization accuracy. RNNs are demonstrated to be effective in modeling the vehicle's dynamics, yielding better temporal smoothness and accuracy. We comprehensively validate the effectiveness of our approach using freshly collected datasets. Multiple trials of repetitive data collection over the same road and areas make our dataset ideal for testing localization systems. The SunnyvaleBigLoop sequences, with a year's time interval between the collected mapping and testing data, made it quite challenging, but the low localization error of our method in these datasets demonstrates its maturity for real industrial implementation.

To understand a scene in depth not only involves locating/recognizing individual objects, but also requires to infer the relationships and interactions among them. However, since the distribution of real-world relationships is seriously unbalanced, existing methods perform quite poorly for the less frequent relationships. In this work, we find that the statistical correlations between object pairs and their relationships can effectively regularize semantic space and make prediction less ambiguous, and thus well address the unbalanced distribution issue. To achieve this, we incorporate these statistical correlations into deep neural networks to facilitate scene graph generation by developing a Knowledge-Embedded Routing Network. More specifically, we show that the statistical correlations between objects appearing in images and their relationships, can be explicitly represented by a structured knowledge graph, and a routing mechanism is learned to propagate messages through the graph to explore their interactions. Extensive experiments on the large-scale Visual Genome dataset demonstrate the superiority of the proposed method over current state-of-the-art competitors.

The key challenge of knowledge distillation is to extract general, moderate and sufficient knowledge from a teacher network to guide a student network. In this paper, a novel Instance Relationship Graph (IRG) is proposed for knowledge distillation. It models three kinds of knowledge, including instance features, instance relationships and feature space transformation, while the latter two kinds of knowledge are neglected by previous methods. Firstly, the IRG is constructed to model the distilled knowledge of one network layer, by considering instance features and instance relationships as vertexes and edges respectively. Secondly, an IRG transformation is proposed to models the feature space transformation across layers. It is more moderate than directly mimicking the features at intermediate layers. Finally, hint loss functions are designed to force a student's IRGs to mimic the structures of a teacher's IRGs. The proposed method effectively captures the knowledge along the whole network via IRGs, and thus shows stable convergence and strong robustness to different network architectures. In addition, the proposed method shows superior performance over existing methods on datasets of various scales.

Both accuracy and efficiency are of significant importance to the task of semantic segmentation. Existing deep FCNs suffer from heavy computations due to a series of high-resolution feature maps for preserving the detailed knowledge in dense estimation. Although reducing the feature map resolution (i.e., applying a large overall stride) via subsampling operations (e.g., polling and convolution striding) can instantly increase the efficiency, it dramatically decreases the estimation accuracy. To tackle this dilemma, we propose a knowledge distillation method tailored for semantic segmentation to improve the performance of the compact FCNs with large overall stride. To handle the inconsistency between the features of the student and teacher network, we optimize the feature similarity in a transferred latent domain formulated by utilizing a pre-trained autoencoder. Moreover, an affinity distillation module is proposed to capture the long-range dependency by calculating the non local interactions across the whole image. To validate the effectiveness of our proposed method, extensive experiments have been conducted on three popular benchmarks: Pascal VOC, Cityscapes and Pascal Context. Built upon a highly competitive baseline, our proposed method can improve the performance of a student network by 2.5% (mIOU boosts from 70.2 to 72.7 on the cityscapes test set) and can train a better compact model with only 8% float operations (FLOPS) of a model that achieves comparable performances.

Neural architectures able to generate variable-length outputs are extremely effective for applications like Machine Translation and Image Captioning. In this paper, we study the vulnerability of these models to attacks aimed at changing the output-size that can have undesirable consequences including increased computation and inducing faults in downstream modules that expect outputs of a certain length. We show the existence and construction of such attacks with two key contributions. First, to overcome the difficulties of discrete search space and the non-differentiable adversarial objective function, we develop an easy-to-compute differentiable proxy objective that can be used with gradient-based algorithms to find output-lengthening inputs. Second, we develop a verification approach to formally prove that the network cannot produce outputs greater than a certain length. Experimental results on Machine Translation and Image Captioning models show that our adversarial output-lengthening approach can produce outputs that are 50 times longer than the input, while our verification approach can, given a model and input domain, prove that the output length is below a certain size.

Machine Learning (ML) models are increasingly deployed in the wild to perform a wide range of tasks. In this work, we ask to what extent can an adversary steal functionality of such "victim" models based solely on blackbox interactions: image in, predictions out. In contrast to prior work, we study complex victim blackbox models, and an adversary lacking knowledge of train/test data used by the model, its internals, and semantics over model outputs. We formulate model functionality stealing as a two-step approach: (i) querying a set of input images to the blackbox model to obtain predictions; and (ii) training a "knockoff" with queried image-prediction pairs. We make multiple remarkable observations: (a) querying random images from a different distribution than that of the blackbox training data results in a well-performing knockoff; (b) this is possible even when the knockoff is represented using a different architecture; and (c) our reinforcement learning approach additionally improves query sample efficiency in certain settings and provides performance gains. We validate model functionality stealing on a range of datasets and tasks, as well as show that a reasonable knockoff of an image analysis API could be created for as little as 30.

Ideally, 360deg imagery could inherit the deep convolutional neural networks (CNNs) already trained with great success on perspective projection images. However, existing methods to transfer CNNs from perspective to spherical images introduce significant computational costs and/or degradations in accuracy. We present the Kernel Transformer Network (KTN) to efficiently transfer convolution kernels from perspective images to the equirectangular projection of 360deg images. Given a source CNN for perspective images as input, the KTN produces a function parameterized by a polar angle and kernel as output. Given a novel 360deg image, that function in turn can compute convolutions for arbitrary layers and kernels as would the source CNN on the corresponding tangent plane projections. Distinct from all existing methods, KTNs allow model transfer: the same model can be applied to different source CNNs with the same base architecture. This enables application to multiple recognition tasks without re-training the KTN. Validating our approach with multiple source CNNs and datasets, we show that KTNs improve the state of the art for spherical convolution. KTNs successfully preserve the source CNN's accuracy, while offering transferability, scalability to typical image resolutions, and, in many cases, a substantially lower memory footprint.

In recent years, scene parsing has captured increasing attention in computer vision. Previous works have demonstrated promising performance in this task. However, they mainly utilize holistic features, whilst neglecting the rich semantic knowledge and inter-object relationships in the scene. In addition, these methods usually require a large number of pixel-level annotations, which is too expensive in practice. In this paper, we propose a novel Knowledge Embedded Generative Adversarial Networks, dubbed as KE-GAN, to tackle the challenging problem in a semi-supervised fashion. KE-GAN captures semantic consistencies of different categories by devising a Knowledge Graph from the large-scale text corpus. In addition to readily-available unlabeled data, we generate synthetic images to unveil rich structural information underlying the images. Moreover, a pyramid architecture is incorporated into the discriminator to acquire multi-scale contextual information for better parsing results. Extensive experimental results on four standard benchmarks demonstrate that KE-GAN is capable of improving semantic consistencies and learning better representations for scene parsing, resulting in the state-of-the-art performance.

Hashing based approximate nearest neighbor search embeds high dimensional data to compact binary codes, which enables efficient similarity search and storage. However, the non-isometry sign() function makes it hard to project the nearest neighbors in continuous data space into the closest codewords in discrete Hamming space. In this work, we revisit the sign() function from the perspective of space partitioning. In specific, we bridge the gap between k-nearest neighbors and binary hashing codes with Shannon entropy. We further propose a novel K-Nearest Neighbors Hashing (KNNH) method to learn binary representations from KNN within the subspaces generated by sign(). Theoretical and experimental results show that the KNN relation is of central importance to neighbor preserving embeddings, and the proposed method outperforms the state-of-the-arts on benchmark datasets.

Given dense image feature correspondences of a non-rigidly moving object across multiple frames, this paper proposes an algorithm to estimate its 3D shape for each frame. To solve this problem accurately, the recent state-of-the-art algorithm reduces this task to set of local linear subspace reconstruction and clustering problem using Grassmann manifold representation [34]. Unfortunately, their method missed on some of the critical issues associated with the modeling of surface deformations, for e.g., the dependence of a local surface deformation on its neighbors. Furthermore, their representation to group high dimensional data points inevitably introduce the drawbacks of categorizing samples on the high-dimensional Grassmann manifold [32, 31]. Hence, to deal with such limitations with [34], we propose an algorithm that jointly exploits the benefit of high-dimensional Grassmann manifold to perform reconstruction, and its equivalent lower-dimensional representation to infer suitable clusters. To accomplish this, we project each Grassmannians onto a lower-dimensional Grassmann manifold which preserves and respects the deformation of the structure w.r.t its neighbors. These Grassmann points in the lower-dimension then act as a representative for the selection of high-dimensional Grassmann samples to perform each local reconstruction. In practice, our algorithm provides a geometrically efficient way to solve dense NRSfM by switching between manifolds based on its benefit and usage. Experimental results show that the proposed algorithm is very effective in handling noise with reconstruction accuracy as good as or better than the competing methods.

In this paper, we propose a semi-supervised approach for representative selection, which finds a small set of representatives that can well summarize a large data collection. Given labeled source data and big unlabeled target data, we aim to find representatives in the target data, which can not only represent and associate data points belonging to each labeled category, but also discover novel categories in the target data, if any. To leverage labeled source data, we guide representative selection from labeled source to unlabeled target. We propose a joint optimization framework which alternately optimizes (1) representative selection in the target data and (2) discriminative feature learning from both the source and the target for better representative selection. Experiments on image and video datasets demonstrate that our proposed approach not only finds better representatives, but also can discover novel categories in the target data that are not in the source.

Facial action unit (AU) intensity is an index to characterize human expressions. Accurate AU intensity estimation depends on three major elements: image representation, intensity estimator, and supervisory information. Most existing methods learn intensity estimator with fixed image representation, and rely on the availability of fully annotated supervisory information. In this paper, a novel general framework for AU intensity estimation is presented, which differs from traditional estimation methods in two aspects. First, rather than keeping image representation fixed, it simultaneously learns representation and intensity estimator to achieve an optimal solution. Second, it allows incorporating weak supervisory training signal from human knowledge (e.g. feature smoothness, label smoothness, label ranking, and positive label), which makes our model trainable even fully annotated information is not available. More specifically, human knowledge is represented as either soft or hard constraints which are encoded as regularization terms or equality/inequality constraints, respectively. On top of our novel framework, we additionally propose an efficient algorithm for optimization based on Alternating Direction Method of Multipliers (ADMM). Evaluations on two benchmark databases show that our method outperforms competing methods under different ratios of AU intensity annotations, especially for small ratios.

We present a new predictor combination algorithm that improves a given task predictor based on potentially relevant reference predictors. Existing approaches are limited in that, to discover the underlying task dependence, they either require known parametric forms of all predictors or access to a single fixed dataset on which all predictors are jointly evaluated. To overcome these limitations, we design a new non-parametric task dependence estimation procedure that automatically aligns evaluations of heterogeneous predictors across disjoint feature sets. Our algorithm is instantiated as a robust manifold diffusion process that jointly refines the estimated predictor alignments and the corresponding task dependence. We apply this algorithm to the relative attributes ranking problem and demonstrate that it not only broadens the application range of predictor combination approaches but also outperforms existing methods even when applied to classical predictor combination settings.

Facial motion retargeting is an important problem in both computer graphics and vision, which involves capturing the performance of a human face and transferring it to another 3D character. Learning 3D morphable model (3DMM) parameters from 2D face images using convolutional neural networks is common in 2D face alignment, 3D face reconstruction etc. However, existing methods either require an additional face detection step before retargeting or use a cascade of separate networks to perform detection followed by retargeting in a sequence. In this paper, we present a single end-to-end network to jointly predict the bounding box locations and 3DMM parameters for multiple faces. First, we design a novel multitask learning framework that learns a disentangled representation of 3DMM parameters for a single face. Then, we leverage the trained single face model to generate ground truth 3DMM parameters for multiple faces to train another network that performs joint face detection and motion retargeting for images with multiple faces. Experimental results show that our joint detection and retargeting network has high face detection accuracy and is robust to extreme expressions and poses while being faster than state-of-the-art methods.

Deep learning approaches to optical flow estimation have seen rapid progress over the recent years. One common trait of many networks is that they refine an initial flow estimate either through multiple stages or across the levels of a coarse-to-fine representation. While leading to more accurate results, the downside of this is an increased number of parameters. Taking inspiration from both classical energy minimization approaches as well as residual networks, we propose an iterative residual refinement (IRR) scheme based on weight sharing that can be combined with several backbone networks. It reduces the number of parameters, improves the accuracy, or even achieves both. Moreover, we show that integrating occlusion prediction and bi-directional flow estimation into our IRR scheme can further boost the accuracy. Our full network achieves state-of-the-art results for both optical flow and occlusion estimation across several standard datasets.