ICCV 2017 Papers

Deep convolutional neural networks have achieved significant improvements on face recognition task due to their ability to learn highly discriminative features from tremendous amounts of face images. Many large scale face datasets exhibit long-tail distribution where a small number of entities (persons) have large number of face images while a large number of persons only have very few face samples (long tail). Most of the existing works alleviate this problem by simply cutting the tailed data and only keep identities with enough number of examples. Unlike these work, this paper investigated how long-tailed data impact the training of face CNNs and develop a novel loss function, called range loss, to effectively utilize the tailed data in training process. More specifically, range loss is designed to reduce overall intrapersonal variations while enlarge inter-personal differences simultaneously. Extensive experiments on two face recognition benchmarks, Labeled Faces in the Wild (LFW) and YouTube Faces (YTF), demonstrate the effectiveness of the proposed range loss in overcoming the long tail effect, and show the good generalization ability of the proposed methods.

Multi-person pose estimation in the wild is challenging. Although state-of-the-art human detectors have demonstrated good performance, small errors in localization and recognition are inevitable. These errors can cause failures for a single-person pose estimator (SPPE), especially for methods that solely depend on human detection results. In this paper, we propose a novel regional multi-person pose estimation (RMPE) framework to facilitate pose estimation in the presence of inaccurate human bounding boxes. Our framework consists of three components: Symmetric Spatial Transformer Network (SSTN), Parametric Pose Non-Maximum-Suppression (NMS), and Pose-Guided Proposals Generator (PGPG). Our method is able to handle inaccurate bounding boxes and redundant detections, allowing it to achieve 76.7 mAP on the MPII (multi person) dataset. Our model and source codes are made publicly available.

Person re-identification (Re-ID) is an important problem in video surveillance, aiming to match pedestrian images across camera views. Currently, most works focus on RGB-based Re-ID. However, in some applications, RGB images are not suitable, e.g. in a dark environment or at night. Infrared (IR) imaging becomes necessary in many visual systems. To that end, matching RGB images with infrared images is required, which are heterogeneous with very different visual characteristics. For person Re-ID, this is a very challenging cross-modality problem that has not been studied so far. In this work, we address the RGB-IR cross-modality Re-ID problem and contribute a new multiple modality Re-ID dataset named SYSU-MM01, including RGB and IR images of 491 identities from 6 cameras, giving in total 287,628 RGB images and 15,792 IR images. To explore the RGB-IR Re-ID problem, we evaluate existing popular cross-domain models, including three commonly used neural network structures (one-stream, two-stream and asymmetric FC layer) and analyse the relation between them. We further propose deep zero-padding for training one-stream network towards automatically evolving domain-specific nodes in the network for cross-modality matching. Our experiments show that RGB-IR cross-modality matching is very challenging but still feasible using the proposed model with deep zero-padding, giving the best performance. Our dataset is available at http://isee.sysu.edu.cn/project/RGBIRReID.htm.

In multi-class indoor semantic segmentation using RGB-D data, it has been shown that incorporating depth feature into RGB feature is helpful to improve segmentation accuracy. However, previous studies have not fully exploited the potentials of multi-modal feature fusion, e.g., simply concatenating RGB and depth features or averaging RGB and depth score maps. To learn the optimal fusion of multi-modal features, this paper presents a novel network that extends the core idea of residual learning to RGB-D semantic segmentation. Our network effectively captures multi-level RGB-D CNN features by including multi-modal feature fusion blocks and multi-level feature refinement blocks. Feature fusion blocks learn residual RGB and depth features and their combinations to fully exploit the complementary characteristics of RGB and depth data. Feature refinement blocks learn the combination of fused features from multiple levels to enable high-resolution prediction. Our network can efficiently train discriminative multi-level features from each modality end-to-end by taking full advantage of skip-connections. Our comprehensive experiments demonstrate that the proposed architecture achieves the state-of-the-art accuracy on two challenging RGB-D indoor datasets, NYUDv2 and SUN RGB-D.

We address the problem of activity detection in continuous, untrimmed video streams. This is a difficult task that requires extracting meaningful spatio-temporal features to capture activities, accurately localizing the start and end times of each activity. We introduce a new model, Region Convolutional 3D Network (R-C3D), which encodes the video streams using a three-dimensional fully convolutional network, then generates candidate temporal regions containing activities, and finally classifies selected regions into specific activities. Computation is saved due to the sharing of convolutional features between the proposal and the classification pipelines. The entire model is trained end-to-end with jointly optimized localization and classification losses. R-C3D is faster than existing methods (569 frames per second on a single Titan X Maxwell GPU) and achieves state-of-the-art results on THUMOS'14. We further demonstrate that our model is a general activity detection framework that does not rely on assumptions about particular dataset properties by evaluating our approach on ActivityNet and Charades. Our code is available at http://ai.bu.edu/r-c3d/.

Triangulation is a fundamental task in 3D computer vision. Unsurprisingly, it is a well-investigated problem with many mature algorithms. However, algorithms for robust triangulation, which are necessary to produce correct results in the presence of egregiously incorrect measurements (i.e., outliers), have received much less attention. The default approach to deal with outliers in triangulation is by random sampling. The randomized heuristic is not only suboptimal, it could, in fact, be computationally inefficient on large-scale datasets. In this paper, we propose a novel locally optimal algorithm for robust triangulation. A key feature of our method is to efficiently derive the local update step by plane sweeping a set of quasiconvex functions. Underpinning our method is a new theory behind quasiconvex plane sweep, which has not been examined previously in computational geometry. Relative to the random sampling heuristic, our algorithm not only guarantees deterministic convergence to a local minimum, it typically achieves higher quality solutions in similar runtimes.

The power of modern image matching approaches is still fundamentally limited by the abrupt scale changes in images. In this paper, we propose a scale-invariant image matching approach to tackling the very large scale variation of views. Drawing inspiration from the scale space theory, we start with encoding the image's scale space into a compact multi-scale representation. Then, rather than trying to find the exact feature matches all in one step, we propose a progressive two-stage approach. First, we determine the related scale levels in scale space, enclosing the inlier feature correspondences, based on an optimal and exhaustive matching in a limited scale space. Second, we produce both the image similarity measurement and feature correspondences simultaneously after restricting matching between the related scale levels in a robust way. The matching performance has been intensively evaluated on vision tasks including image retrieval, feature matching and Structure-from-Motion (SfM). The successful integration of the challenging fusion of high aerial and low ground-level views with significant scale differences manifests the superiority of the proposed approach.

In this paper we deal with the problem of recovering affine camera calibration and objects position/occupancy from multi-view images using the information from image detections. We show that remarkable object localisation and volumetric occupancy can be recovered by including both geometrical constraints and prior information given by objects CAD models from the ShapeNet dataset. This can be done by recasting the problem in the context of a probabilistic framework based on Probabilistic PCA that includes both the object semantic priors together with the multi-view geometrical constraints. We present results on synthetic and real datasets to show the validity of our approach and improvements with respect to previous approaches. In particular, the statistical priors are key to obtain reliable 3D reconstruction especially when the input detections are noisy, a likely case in real scenarios.

Optical flow estimation remains challenging due to untextured areas, motion boundaries, occlusions, and more. Thus, the estimated flow is not equally reliable across the image. To that end, post-hoc confidence measures have been introduced to assess the per-pixel reliability of the flow. We overcome the artificial separation of optical flow and confidence estimation by introducing a method that jointly predicts optical flow and its underlying uncertainty. Starting from common energy-based formulations, we rely on the corresponding posterior distribution of the flow given the images. We derive a variational inference scheme based on mean field, which incorporates best practices from energy minimization. An uncertainty measure is obtained along the flow at every pixel as the (marginal) entropy of the variational distribution. We demonstrate the flexibility of our probabilistic approach by applying it to two different energies and on two benchmarks. We not only obtain flow results that are competitive with the underlying energy minimization approach, but also a reliable uncertainty measure that significantly outperforms existing post-hoc approaches.

We propose a privacy-preserving framework for learning visual classifiers by leveraging distributed private image data. This framework is designed to aggregate multiple classifiers updated locally using private data and to ensure that no private information about the data is exposed during and after its learning procedure. We utilize a homomorphic cryptosystem that can aggregate the local classifiers while they are encrypted and thus kept secret. To overcome the high computational cost of homomorphic encryption of high-dimensional classifiers, we (1) impose sparsity constraints on local classifier updates and (2) propose a novel efficient encryption scheme named doubly-permuted homomorphic encryption (DPHE) which is tailored to sparse high-dimensional data. DPHE (i) decomposes sparse data into its constituent non-zero values and their corresponding support indices, (ii) applies homomorphic encryption only to the non-zero values, and (iii) employs double permutations on the support indices to make them secret. Our experimental evaluation on several public datasets shows that the proposed approach achieves comparable performance against state-of-the-art visual recognition methods while preserving privacy and significantly outperforms other privacy-preserving methods.

This paper proposes a novel approach for segmenting primary video objects by using Complementary Convolutional Neural Networks (CCNN) and neighborhood reversible flow. The proposed approach first pre-trains CCNN on massive images with manually annotated salient objects in an end-to-end manner, and the trained CCNN has two separate branches that simultaneously handle two complementary tasks, i.e., foregroundness and backgroundness estimation. By applying CCNN on each video frame, the spatial foregroundness and backgroundness maps can be initialized, which are then propagated between various frames so as to segment primary video objects and suppress distractors. To enforce efficient temporal propagation, we divide each frame into superpixels and construct neighborhood reversible flow that reflects the most reliable temporal correspondences between superpixels in far-away frames. Within such flow, the initialized foregroundness and backgroundness can be efficiently and accurately propagated along the temporal axis so that primary video objects gradually pop-out and distractors are well suppressed. Extensive experimental results on three video datasets show that the proposed approach achieves impressive performance in comparisons with 18 state-of-the-art models.

We present an algorithm for test-time combination of a set of reference predictors with unknown parametric forms. Existing multi-task and transfer learning algorithms focus on training-time transfer and combination, where the parametric forms of predictors are known and shared. However, when the parametric form of a predictor is unknown, e.g., for a human predictor or a predictor in a precompiled library, existing algorithms are not applicable. Instead, we empirically evaluate predictors on sampled data points to measure distances between different predictors. This embeds the set of reference predictors into a Riemannian manifold, upon which we perform manifold denoising to obtain the refined predictor. This allows our approach to make no assumptions about the underlying predictor forms. Our test-time combination algorithm equals or outperforms existing multi-task and transfer learning algorithms on challenging real-world datasets, without introducing specific model assumptions.

Leveraging class semantic descriptions and examples of known objects, zero-shot learning makes it possible to train a recognition model for an object class whose examples are not available. In this paper, we propose a novel zero-shot learning model that takes advantage of clustering structures in the semantic embedding space. The key idea is to impose the structural constraint that semantic representations must be predictive of the locations of their corresponding visual exemplars. To this end, this reduces to training multiple kernel-based regressors from semantic representation-exemplar pairs from labeled data of the seen object categories. Despite its simplicity, our approach significantly outperforms existing zero-shot learning methods in three out of four benchmark datasets, including the ImageNet dataset with more than 20,000 unseen categories.

This paper presents a novel method to predict future human activities from partially observed RGB-D videos. Human activity prediction is generally difficult due to its non-Markovian property and the rich context between human and environments. We use a stochastic grammar model to capture the compositional structure of events, integrating human actions, objects, and their affordances. We represent the event by a spatial-temporal And-Or graph (ST-AOG). The ST-AOG is composed of a temporal stochastic grammar defined on sub-activities, and spatial graphs representing sub-activities that consist of human actions, objects, and their affordances. Future sub-activities are predicted using the temporal grammar and Earley parsing algorithm. The corresponding action, object, and affordance labels are then inferred accordingly. Extensive experiments are conducted to show the effectiveness of our model on both semantic event parsing and future activity prediction.

The ability to predict and therefore to anticipate the future is an important attribute of intelligence. It is also of utmost importance in real-time systems, e.g . in robotics or autonomous driving, which depend on visual scene understanding for decision making. While prediction of the raw RGB pixel values in future video frames has been studied in previous work, here we introduce the novel task of predicting semantic segmentations of future frames. Given a sequence of video frames, our goal is to predict segmentation maps of not yet observed video frames that lie up to a second or further in the future. We develop an autoregressive convolutional neural network that learns to iteratively generate multiple frames. Our results on the Cityscapes dataset show that directly predicting future segmentations is substantially better than predicting and then segmenting future RGB frames. Prediction results up to half a second in the future are visually convincing and are much more accurate than those of a baseline based on warping semantic segmentations using optical flow.

Face alignment has witnessed substantial progress in the last decade. One of the recent focuses has been aligning a dense 3D face shape to face images with large head poses. The dominant technology used is based on the cascade of regressors, e.g., CNNs, which has shown promising results. Nonetheless, the cascade of CNNs suffers from several drawbacks, e.g., lack of end-to-end training, hand-crafted features and slow training speed. To address these issues, we propose a new layer, named visualization layer, which can be integrated into the CNN architecture and enables joint optimization with different loss functions. Extensive evaluation of the proposed method on multiple datasets demonstrates state-of-the-art accuracy, while reducing the training time by more than half compared to the typical cascade of CNNs. In addition, we compare across multiple CNN architectures, all with the visualization layer, to further demonstrate the advantage of its utilization.

Feature extraction and matching are two crucial components in person Re-Identification (ReID). The large pose deformations and the complex view variations exhibited by the captured person images significantly increase the difficulty of learning and matching of the features from person images. To overcome these difficulties, in this work we propose a Pose-driven Deep Convolutional (PDC) model to learn improved feature extraction and matching models from end to end. Our deep architecture explicitly leverages the human part cues to alleviate the pose variations and learn robust feature representations from both the global image and different local parts. To match the features from global human body and local body parts, a pose driven feature weighting sub-network is further designed to learn adaptive feature fusions. Extensive experimental analyses and results on three popular datasets demonstrate significant performance improvements of our model over all published stateof- the-art methods.

In this paper we examine the effects of using object poses as guidance to learning robust features for 3D object pose estimation. Previous works have focused on learning feature embeddings based on metric learning with triplet comparisons and rely only on the qualitative distinction of similar and dissimilar pose labels. In contrast, we consider the exact pose differences between the training samples, and aim to learn embeddings such that the distances in the pose label space are proportional to the distances in the feature space. However, since it is less desirable to force the pose-feature correlation when objects are symmetric, we propose the data-driven weights that reflect object symmetry when measuring the pose distances. Furthermore, end-to-end pose regression is investigated and is shown to further boost the discriminative power of feature learning, improving pose recognition accuracies in NN, and thus can be used as another pose guidance to feature learning. Experimental results show that the features guided by poses, are significantly more discriminative than the ones learned in the traditional way, outperforming state-of-the-art works. Finally, we measure the generalisation capacities of pose guided feature learning in previously unseen scenes containing objects under different occlusion levels, and we show that it adapts well to novel tasks.

In this paper we present a new method for creating polynomial solvers for problems where a (possibly infinite) subset of the solutions are undesirable or uninteresting. These solutions typically arise from simplifications made during modeling, but can also come from degeneracies which are inherent to the geometry of the original problem. The proposed approach extends the standard action matrix method to saturated ideals. This allows us to add constraints that some polynomials should be non-zero on the solutions. This does not only offer the possibility of improved performance by removing superfluous solutions, but makes a larger class of problems tractable. Previously, problems with infinitely many solutions could not be solved directly using the action matrix method as it requires a zero-dimensional ideal. In contrast we only require that after removing the unwanted solutions only finitely many remain. We evaluate our method on three applications, optimal triangulation, time-of-arrival self-calibration and optimal vanishing point estimation.