ICCV 2017 Papers

We present an end-to-end system for detecting and clustering faces by identity in full-length movies. Unlike works that start with a predefined set of detected faces, we consider the end-to-end problem of detection and clustering together. We make three separate contributions. First, we combine a state-of-the-art face detector with a generic tracker to extract high quality face tracklets. We then introduce a novel clustering method, motivated by the classic graph theory results of Erdos and Renyi. It is based on the observations that large clusters can be fully connected by joining just a small fraction of their point pairs, while just a single connection between two different people can lead to poor clustering results. This suggests clustering using a verification system with very few false positives but perhaps moderate recall. We introduce a novel verification method, rank-1 counts verification, that has this property, and use it in a link-based clustering scheme. Finally, we define a novel end-to-end detection and clustering evaluation metric allowing us to assess the accuracy of the entire end-to-end system. We present state-of-the-art results on multiple video data sets and also on standard face databases.

In contrast to the widely studied problem of recognizing an action given a complete sequence, action anticipation aims to identify the action from only partially available videos. As such, it is therefore key to the success of computer vision applications requiring to react as early as possible, such as autonomous navigation. In this paper, we propose a new action anticipation method that achieves high prediction accuracy even in the presence of a very small percentage of a video sequence. To this end, we develop a multi-stage LSTM architecture that leverages context-aware and action-aware features, and introduce a novel loss function that encourages the model to predict the correct class as early as possible. Our experiments on standard benchmark datasets evidence the benefits of our approach; We outperform the state-of-the-art action anticipation methods for early prediction by a relative increase in accuracy of 22.0% on JHMDB-21, 14.0% on UT-Interaction and 49.9% on UCF-101.

This paper introduces a new lifelong learning solution where a single model is trained for a sequence of tasks. The main challenge that vision systems face in this context is catastrophic forgetting: as they tend to adapt to the most recently seen task, they lose performance on the tasks that were learned previously. Our method aims at preserving the knowledge of the previous tasks while learning a new one by using autoencoders. For each task, an under-complete autoencoder is learned, capturing the features that are crucial for its achievement. When a new task is presented to the system, we prevent the reconstructions of the features with these autoencoders from changing, which has the effect of preserving the information on which the previous tasks are mainly relying. At the same time, the features are given space to adjust to the most recent environment as only their projection into a low dimension submanifold is controlled. The proposed system is evaluated on image classification tasks and shows a reduction of forgetting over the state-of-the-art.

Object part segmentation is a challenging and fundamental problem in computer vision. Its difficulties may be caused by the varying viewpoints, poses, and topological structures, which can be attributed to an essential reason, i.e., a specific object is a 3D model rather than a 2D figure. Therefore, we conjecture that not only 2D appearance features but also 3D geometric features could be helpful. With this in mind, we propose a 2-stream FCN. One stream, named AppNet, is to extract 2D appearance features from the input image. The other stream, named GeoNet, is to extract 3D geometric features. However, the problem is that the input is just an image. To this end, we design a 2D-convolution based CNN structure to extract 3D geometric features from 3D volume, which is named VolNet. Then a teacher-student strategy is adopted and VolNet teaches GeoNet how to extract 3D geometric features from an image. To perform this teaching process, we synthesize training data using 3D models. Each training sample consists of an image and its corresponding volume. A perspective voxelization algorithm is further proposed to align them. Experimental results verify our conjecture and the effectiveness of both the proposed 2-stream CNN and VolNet.

Gesture is a natural interface in interacting with wearable devices such as VR/AR helmet and glasses. The main challenge of gesture recognition in egocentric vision arises from the global camera motion caused by the spontaneous head movement of the device wearer. In this paper, we address the problem by a novel recurrent 3D convolutional neural network for end-to-end learning. We specially design a spatiotemporal transformer module with recurrent connections between neighboring time slices which can actively transform a 3D feature map into a canonical view in both spatial and temporal dimensions. To validate our method, we introduce a new dataset with sufficient size, variation and reality, which contains 83 gestures designed for interaction with wearable devices, and more than 24,000 RGB-D gesture samples from 50 subjects captured in 6 scenes. On this dataset, we show that the proposed network outperforms competing state-of-the-art algorithms. Moreover, our method can achieve state-of-the-art performance on the challenging GTEA egocentric action dataset.

Many existing person re-identification (PRID) methods typically attempt to train a faithful global metric offline to cover the enormous visual appearance variations, so as to directly use it online on various probes for identity matching. However, their need for a huge set of positive training pairs is very demanding in practice. In contrast to these methods, this paper advocates a different paradigm: part of the learning can be performed online but with nominal costs, so as to achieve online metric adaptation for different input probes. A major challenge here is that no positive training pairs are available for the probe anymore. By only exploiting easily-available negative samples, we propose a novel solution to achieve local metric adaptation effectively and efficiently. For each probe at the test time, it learns a strictly positive semi-definite dedicated local metric. Comparing to offline global metric learning, its computational cost is negligible. The insight of this new method is that the local hard negative samples can actually provide tight constraints to fine tune the metric locally. This new local metric adaptation method is generally applicable, as it can be used on top of any global metric to enhance its performance. In addition, this paper gives in-depth theoretical analysis and justification of the new method. We prove that our new method guarantees the reduction of the classification error asymptotically, and prove that it actually learns the optimal local metric to best approximate the asymptotic case by a finite number of training data. Extensive experiments and comparative studies on almost all major benchmarks (VIPeR, QMUL GRID, CUHK Campus, CUHK03 and Market-1501) have confirmed the effectiveness and superiority of our method.

Using the matrix product state (MPS) representation of the recently proposed tensor ring (TR) decompositions, in this paper we propose a TR completion algorithm, which is an alternating minimization algorithm that alternates over the factors in the MPS representation. This development is motivated in part by the success of matrix completion algorithms that alternate over the (low-rank) factors. We propose a novel initialization method and analyze the computational complexity of the TR completion algorithm. The numerical comparison between the TR completion algorithm and the existing algorithms that employ a low rank tensor train (TT) approximation for data completion shows that our method outperforms the existing ones for a variety of real computer vision settings, and thus demonstrates the improved expressive power of tensor ring as compared to tensor train.

We propose an efficient self-shadowing illumination model for Morphable Models. Simulating self-shadowing with ray casting is computationally expensive which makes them impractical in Analysis-by-Synthesis methods for object reconstruction from single images. Therefore, we propose to learn self-shadowing for Morphable Model parameters directly with a linear model. Radiance transfer functions are a powerful way to represent self-shadowing used within the precomputed radiance transfer framework (PRT). We build on PRT to render deforming objects with self-shadowing at interactive frame rates. It can be illuminated efficiently by environment maps represented with spherical harmonics. The result is an efficient global illumination method for Morphable Models, exploiting an approximated radiance transfer. We apply the method to fitting Morphable Model parameters to a single image of a face and demonstrate that considering self-shadowing improves shape reconstruction.

Given an image of a street scene in a city, this paper develops a new method that can quickly and precisely pinpoint at which location (as well as viewing direction) the image was taken, against a pre-stored large-scale 3D point-cloud map of the city. We adopt the recently developed 2D-3D direct feature matching framework for this task [23,31,32,42-44]. This is a challenging task especially for large-scale problems. As the map size grows bigger, many 3D points in the wider geographical area can be visually very similar-or even identical-causing severe ambiguities in 2D-3D feature matching. The key is to quickly and unambiguously find the correct matches between a query image and the large 3D map. Existing methods solve this problem mainly via comparing individual features' visual similarities in a local and per feature manner, thus only local solutions can be found, inadequate for large-scale applications. In this paper, we introduce a global method which harnesses global contextual information exhibited both within the query image and among all the 3D points in the map. This is achieved by a novel global ranking algorithm, applied to a Markov network built upon the 3D map, which takes account of not only visual similarities between individual 2D-3D matches, but also their global compatibilities (as measured by co-visibility) among all matching pairs found in the scene. Tests on standard benchmark datasets show that our method achieved both higher precision and comparable recall, compared with the state-of-the-art.

Lineage tracing, the joint segmentation and tracking of living cells as they move and divide in a sequence of light microscopy images, is a challenging task. Jug et al. have proposed a mathematical abstraction of this task, the moral lineage tracing problem (MLTP), whose feasible solutions define both a segmentation of every image and a lineage forest of cells. Their branch-and-cut algorithm, however, is prone to many cuts and slow convergence for large instances. To address this problem, we make three contributions: (i) we devise the first efficient primal feasible local search algorithms for the MLTP, (ii) we improve the branch-and-cut algorithm by separating tighter cutting planes and by incorporating our primal algorithms, (iii) we show in experiments that our algorithms find accurate solutions on the problem instances of Jug et al. and scale to larger instances, leveraging moral lineage tracing to practical significance.

We present an algorithm to compute parametric models of dense foliage. The guiding principles of our work are automatic reconstruction and compact artist friendly representation. We use Bezier patches to model leaf surface, which we compute from images and point clouds of dense foliage. We present an algorithm to segment individual leaves from colour and depth data. We then reconstruct the Bezier representation from segmented leaf points clouds using non-linear optimisation. Unlike previous work, we do not require laboratory scanned exemplars or user intervention. We also demonstrate intuitive manipulators to edit the reconstructed parametric models. We believe our work is a step towards making captured data more accessible to artists for foliage modelling.

This paper considers the multi-camera motion segmentation problem using unsynchronized videos. Specifically, given two video clips containing several moving objects, captured by unregistered, unsynchronized cameras with different viewpoints, our goal is to assign features to moving objects in the scene. This problem challenges existing methods, due to the lack of registration information and correspondences across cameras. To solve it, we propose a new method that exploits both shape and dynamical information and does not require spatio-temporal registration or shared features. As shown in the paper, the combination of shape and dynamical information results in improved performance even in the single camera case, and allows for solving the multi-camera segmentation problem with a computational cost similar to that of existing single-view techniques. These results are illustrated using both the existing Hopkins 155 data set and a new multi-camera data set, the RSL-12.

Label estimation is an important component in an unsupervised person re-identification (re-ID) system. This paper focuses on cross-camera label estimation, which can be subsequently used in feature learning to learn robust re-ID models. Specifically, we propose to construct a graph for samples in each camera, and then graph matching scheme is introduced for cross-camera labeling association. While labels directly output from existing graph matching methods may be noisy and inaccurate due to significant cross-camera variations, this paper propose a dynamic graph matching (DGM) method. DGM iteratively updates the image graph and the label estimation process by learning a better feature space with intermediate estimated labels. DGM is advantageous in two aspects: 1) the accuracy of estimated labels is improved significantly with the iterations; 2) DGM is robust to noisy initial training data. Extensive experiments conducted on three benchmarks including the large-scale MARS dataset show that DGM yields competitive performance to fully supervised baselines, and outperforms competing unsupervised learning methods.

In this work we propose a novel framework named DualNet aiming at learning more accurate representation for image recognition. Here two parallel neural networks are coordinated to learn complementary features and thus a wider network is constructed. Specifically, we logically divide an end-to-end deep convolutional neural network into two functional parts, i.e., feature extractor and image classifier. The extractors of two subnetworks are placed side by side, which exactly form the feature extractor of DualNet. Then the two-stream features are aggregated to the final classifier for overall classification, while two auxiliary classifiers are appended behind the feature extractor of each subnetwork to make the separately learned features discriminative alone. The complementary constraint is imposed by weighting the three classifiers, which is indeed the key of DualNet. The corresponding training strategy is also proposed, consisting of iterative training and joint finetuning, to make the two subnetworks cooperate well with each other. Finally, DualNet based on the well-known CaffeNet, VGGNet, NIN and ResNet are thoroughly investigated and experimentally evaluated on multiple datasets including CIFAR-100, Stanford Dogs and UEC FOOD-100. The results demonstrate that DualNet can really help learn more accurate image representation, and thus result in higher accuracy for recognition. In particular, the performance on CIFAR-100 is state-of-the-art compared to the recent works.

Conditional Generative Adversarial Networks (GANs) for cross-domain image-to-image translation have made much progress recently. Depending on the task complexity, thousands to millions of labeled image pairs are needed to train a conditional GAN. However, human labeling is expensive, even impractical, and large quantities of data may not always be available. Inspired by dual learning from natural language translation, we develop a novel mechanism, which enables image translators to be trained from two sets of images from two domains. In our architecture, the primal GAN learns to translate images from domain U to those in domain V, while the dual GAN learns to invert the task. The closed loop made by the primal and dual tasks allows images from either domain to be translated and then reconstructed. Hence a loss function that accounts for the reconstruction error of images can be used to train the translators. Experiments on multiple image translation tasks with unlabeled data show considerable performance gain of DualGAN over a single GAN. For some tasks, DualGAN can even achieve comparable or slightly better results than conditional GAN trained on fully labeled data.

Since the beginning of early civilizations, social relationships derived from each individual fundamentally form the basis of social structure in our daily life. In the computer vision literature, much progress has been made in scene understanding, such as object detection and scene parsing. Recent research focuses on the relationship between objects based on its functionality and geometrical relations. In this work, we aim to study the problem of social relationship recognition, in still images. We have proposed a dual-glance model for social relationship recognition, where the first glance fixates at the individual pair of interest and the second glance deploys attention mechanism to explore contextual cues. We have also collected a new large scale People in Social Context (PISC) dataset, which comprises of 22,670 images and 76,568 annotated samples from 9 types of social relationship. We provide benchmark results on the PISC dataset, and qualitatively demonstrate the efficacy of the proposed model.

Future frame prediction in videos is a promising avenue for unsupervised video representation learning. Video frames are naturally generated by the inherent pixel flows from preceding frames based on the appearance and motion dynamics in the video. However, existing methods focus on directly hallucinating pixel values, resulting in blurry predictions. In this paper, we develop a dual motion Generative Adversarial Net (GAN) architecture, which learns to explicitly enforce future-frame predictions to be consistent with the pixel-wise flows in the video through a dual-learning mechanism. The primal future-frame prediction and dual future-flow prediction form a closed loop, generating informative feedback signals to each other for better video prediction. To make both synthesized future frames and flows indistinguishable from reality, a dual adversarial training method is proposed to ensure that the future-flow prediction is able to help infer realistic future-frames, while the future-frame prediction in turn leads to realistic optical flows. Our dual motion GAN also handles natural motion uncertainty in different pixel locations with a new probabilistic motion encoder, which is based on variational autoencoders. Extensive experiments demonstrate that the proposed dual motion GAN significantly outperforms state-of-the-art approaches on synthesizing new video frames and predicting future flows. Our model generalizes well across diverse visual scenes and shows superiority in unsupervised video representation learning.

Existing counting methods often adopt regression-based approaches and cannot precisely localize the target objects, which hinders the further analysis (e.g., high-level understanding and fine-grained classification). In addition, most of prior work mainly focus on counting objects in static environments with fixed cameras. Motivated by the advent of unmanned flying vehicles (i.e., drones), we are interested in detecting and counting objects in such dynamic environments. We propose Layout Proposal Networks (LPNs) and spatial kernels to simultaneously count and localize target objects (e.g., cars) in videos recorded by the drone. Different from the conventional region proposal methods, we leverage the spatial layout information (e.g., cars often park regularly) and introduce these spatially regularized constraints into our network to improve the localization accuracy. To evaluate our counting method, we present a new large-scale car parking lot dataset (CARPK) that contains nearly 90,000 cars captured from different parking lots. To the best of our knowledge, it is the first and the largest drone view dataset that supports object counting, and provides the bounding box annotations.

Deep Neural Networks trained on large datasets can be easily transferred to new domains with far fewer labeled examples by a process called fine-tuning. This has the advantage that representations learned in the large source domain can be exploited on smaller target domains. However, networks designed to be optimal for the source task are often prohibitively large for the target task. In this work we address the compression of networks after domain transfer. We focus on compression algorithms based on low-rank matrix decomposition. Existing methods base compression solely on learned network weights and ignore the statistics of network activations. We show that domain transfer leads to large shifts in network activations and that it is desirable to take this into account when compressing. We demonstrate that considering activation statistics when compressing weights leads to a rank-constrained regression problem with a closed-form solution. Because our method takes into account the target domain, it can more optimally remove the redundancy in the weights. Experiments show that our Domain Adaptive Low Rank (DALR) method significantly outperforms existing low-rank compression techniques. With our approach, the fc6 layer of VGG19 can be compressed more than 4x more than using truncated SVD alone -- with only a minor or no loss in accuracy. When applied to domain-transferred networks it allows for compression down to only 5-20% of the original number of parameters with only a minor drop in performance.