Research papers and code for "Xiaogang Wang":
In this paper, we propose a fast fully convolutional neural network (FCNN) for crowd segmentation. By replacing the fully connected layers in CNN with 1 by 1 convolution kernels, FCNN takes whole images as inputs and directly outputs segmentation maps by one pass of forward propagation. It has the property of translation invariance like patch-by-patch scanning but with much lower computation cost. Once FCNN is learned, it can process input images of any sizes without warping them to a standard size. These attractive properties make it extendable to other general image segmentation problems. Based on FCNN, a multi-stage deep learning is proposed to integrate appearance and motion cues for crowd segmentation. Both appearance filters and motion filers are pretrained stage-by-stage and then jointly optimized. Different combination methods are investigated. The effectiveness of our approach and component-wise analysis are evaluated on two crowd segmentation datasets created by us, which include image frames from 235 and 11 scenes, respectively. They are currently the largest crowd segmentation datasets and will be released to the public.

* 9 pages,7 figures
Click to Read Paper and Get Code
We propose a framework to understand the unprecedented performance and robustness of deep neural networks using field theory. Correlations between the weights within the same layer can be described by symmetries in that layer, and networks generalize better if such symmetries are broken to reduce the redundancies of the weights. Using a two parameter field theory, we find that the network can break such symmetries itself towards the end of training in a process commonly known in physics as spontaneous symmetry breaking. This corresponds to a network generalizing itself without any user input layers to break the symmetry, but by communication with adjacent layers. In the layer decoupling limit applicable to residual networks (He et al., 2015), we show that the remnant symmetries that survive the non-linear layers are spontaneously broken. The Lagrangian for the non-linear and weight layers together has striking similarities with the one in quantum field theory of a scalar. Using results from quantum field theory we show that our framework is able to explain many experimentally observed phenomena,such as training on random labels with zero error (Zhang et al., 2017), the information bottleneck, the phase transition out of it and gradient variance explosion (Shwartz-Ziv & Tishby, 2017), shattered gradients (Balduzzi et al., 2017), and many more.

Click to Read Paper and Get Code
Markov Chain Monte Carlo (MCMC) sampling methods are widely used but often encounter either slow convergence or biased sampling when applied to multimodal high dimensional distributions. In this paper, we present a general framework of improving classical MCMC samplers by employing a global optimization method. The global optimization method first reduces a high dimensional search to an one dimensional geodesic to find a starting point close to a local mode. The search is accelerated and completed by using a local search method such as BFGS. We modify the target distribution by extracting a local Gaussian distribution aound the found mode. The process is repeated to find all the modes during sampling on the fly. We integrate the optimization algorithm into the Wormhole Hamiltonian Monte Carlo (WHMC) method. Experimental results show that, when applied to high dimensional, multimodal Gaussian mixture models and the network sensor localization problem, the proposed method achieves much faster convergence, with relative error from the mean improved by about an order of magnitude than WHMC in some cases.

Click to Read Paper and Get Code
In this paper, we propose PointRCNN for 3D object detection from raw point cloud. The whole framework is composed of two stages: stage-1 for the bottom-up 3D proposal generation and stage-2 for refining proposals in the canonical coordinates to obtain the final detection results. Instead of generating proposals from RGB image or projecting point cloud to bird's view or voxels as previous methods do, our stage-1 sub-network directly generates a small number of high-quality 3D proposals from point cloud in a bottom-up manner via segmenting the point cloud of whole scene into foreground points and background. The stage-2 sub-network transforms the pooled points of each proposal to canonical coordinates to learn better local spatial features, which is combined with global semantic features of each point learned in stage-1 for accurate box refinement and confidence prediction. Extensive experiments on the 3D detection benchmark of KITTI dataset show that our proposed architecture outperforms state-of-the-art methods with remarkable margins by using only point cloud as input.

Click to Read Paper and Get Code
Despite the great success of two-stage detectors, single-stage detector is still a more elegant and efficient way, yet suffers from the two well-known disharmonies during training, i.e. the huge difference in quantity between positive and negative examples as well as between easy and hard examples. In this work, we first point out that the essential effect of the two disharmonies can be summarized in term of the gradient. Further, we propose a novel gradient harmonizing mechanism (GHM) to be a hedging for the disharmonies. The philosophy behind GHM can be easily embedded into both classification loss function like cross-entropy (CE) and regression loss function like smooth-$L_1$ ($SL_1$) loss. To this end, two novel loss functions called GHM-C and GHM-R are designed to balancing the gradient flow for anchor classification and bounding box refinement, respectively. Ablation study on MS COCO demonstrates that without laborious hyper-parameter tuning, both GHM-C and GHM-R can bring substantial improvement for single-stage detector. Without any whistles and bells, our model achieves 41.6 mAP on COCO test-dev set which surpasses the state-of-the-art method, Focal Loss (FL) + $SL_1$, by 0.8.

* To appear at AAAI 2019
Click to Read Paper and Get Code
We propose a new class of transforms that we call {\it Lehmer Transform} which is motivated by the {\it Lehmer mean function}. The proposed {\it Lehmer transform} decomposes a function of a sample into their constituting statistical moments. Theoretical properties of the proposed transform are presented. This transform could be very useful to provide an alternative method in analyzing non-stationary signals such as brain wave EEG.

Click to Read Paper and Get Code
Feature matters. How to train a deep network to acquire discriminative features across categories and polymerized features within classes has always been at the core of many computer vision tasks, specially for large-scale recognition systems where test identities are unseen during training and the number of classes could be at million scale. In this paper, we address this problem based on the simple intuition that the cosine distance of features in high-dimensional space should be close enough within one class and far away across categories. To this end, we proposed the congenerous cosine (COCO) algorithm to simultaneously optimize the cosine similarity among data. It inherits the softmax property to make inter-class features discriminative as well as shares the idea of class centroid in metric learning. Unlike previous work where the center is a temporal, statistical variable within one mini-batch during training, the formulated centroid is responsible for clustering inner-class features to enforce them polymerized around the network truncus. COCO is bundled with discriminative training and learned end-to-end with stable convergence. Experiments on five benchmarks have been extensively conducted to verify the effectiveness of our approach on both small-scale classification task and large-scale human recognition problem.

Click to Read Paper and Get Code
Person recognition aims at recognizing the same identity across time and space with complicated scenes and similar appearance. In this paper, we propose a novel method to address this task by training a network to obtain robust and representative features. The intuition is that we directly compare and optimize the cosine distance between two features - enlarging inter-class distinction as well as alleviating inner-class variance. We propose a congenerous cosine loss by minimizing the cosine distance between samples and their cluster centroid in a cooperative way. Such a design reduces the complexity and could be implemented via softmax with normalized inputs. Our method also differs from previous work in person recognition that we do not conduct a second training on the test subset. The identity of a person is determined by measuring the similarity from several body regions in the reference set. Experimental results show that the proposed approach achieves better classification accuracy against previous state-of-the-arts.

* Post-rebuttal update. Add some comparison results; correct some technical part; rewrite some sections to make it more readable; code link available
Click to Read Paper and Get Code
The recent successful deep neural networks are largely trained in a supervised manner. It {\it associates} complex patterns of input samples with neurons in the last layer, which form representations of {\it concepts}. In spite of their successes, the properties of complex patterns associated a learned concept remain elusive. In this work, by analyzing how neurons are associated with concepts in supervised networks, we hypothesize that with proper priors to regulate learning, neural networks can automatically associate neurons in the intermediate layers with concepts that are aligned with real world concepts, when trained only with labels that associate concepts with top level neurons, which is a plausible way for unsupervised learning. We develop a prior to verify the hypothesis and experimentally find the proposed prior help neural networks automatically learn both basic physical concepts at the lower layers, e.g., rotation of filters, and highly semantic concepts at the higher layers, e.g., fine-grained categories of an entry-level category.

Click to Read Paper and Get Code
As a widely used non-linear activation, Rectified Linear Unit (ReLU) separates noise and signal in a feature map by learning a threshold or bias. However, we argue that the classification of noise and signal not only depends on the magnitude of responses, but also the context of how the feature responses would be used to detect more abstract patterns in higher layers. In order to output multiple response maps with magnitude in different ranges for a particular visual pattern, existing networks employing ReLU and its variants have to learn a large number of redundant filters. In this paper, we propose a multi-bias non-linear activation (MBA) layer to explore the information hidden in the magnitudes of responses. It is placed after the convolution layer to decouple the responses to a convolution kernel into multiple maps by multi-thresholding magnitudes, thus generating more patterns in the feature space at a low computational cost. It provides great flexibility of selecting responses to different visual patterns in different magnitude ranges to form rich representations in higher layers. Such a simple and yet effective scheme achieves the state-of-the-art performance on several benchmarks.

Click to Read Paper and Get Code
In existing works that learn representation for object detection, the relationship between a candidate window and the ground truth bounding box of an object is simplified by thresholding their overlap. This paper shows information loss in this simplification and picks up the relative location/size information discarded by thresholding. We propose a representation learning pipeline to use the relationship as supervision for improving the learned representation in object detection. Such relationship is not limited to object of the target category, but also includes surrounding objects of other categories. We show that image regions with multiple contexts and multiple rotations are effective in capturing such relationship during the representation learning process and in handling the semantic and visual variation caused by different window-object configurations. Experimental results show that the representation learned by our approach can improve the object detection accuracy by 6.4% in mean average precision (mAP) on ILSVRC2014. On the challenging ILSVRC2014 test dataset, 48.6% mAP is achieved by our single model and it is the best among published results. On PASCAL VOC, it outperforms the state-of-the-art result of Fast RCNN by 3.3% in absolute mAP.

* 9 pages, including 1 reference page, 6 figures
Click to Read Paper and Get Code
This paper proposes to learn high-performance deep ConvNets with sparse neural connections, referred to as sparse ConvNets, for face recognition. The sparse ConvNets are learned in an iterative way, each time one additional layer is sparsified and the entire model is re-trained given the initial weights learned in previous iterations. One important finding is that directly training the sparse ConvNet from scratch failed to find good solutions for face recognition, while using a previously learned denser model to properly initialize a sparser model is critical to continue learning effective features for face recognition. This paper also proposes a new neural correlation-based weight selection criterion and empirically verifies its effectiveness in selecting informative connections from previously learned models in each iteration. When taking a moderately sparse structure (26%-76% of weights in the dense model), the proposed sparse ConvNet model significantly improves the face recognition performance of the previous state-of-the-art DeepID2+ models given the same training data, while it keeps the performance of the baseline model with only 12% of the original parameters.

Click to Read Paper and Get Code
We present highly efficient algorithms for performing forward and backward propagation of Convolutional Neural Network (CNN) for pixelwise classification on images. For pixelwise classification tasks, such as image segmentation and object detection, surrounding image patches are fed into CNN for predicting the classes of centered pixels via forward propagation and for updating CNN parameters via backward propagation. However, forward and backward propagation was originally designed for whole-image classification. Directly applying it to pixelwise classification in a patch-by-patch scanning manner is extremely inefficient, because surrounding patches of pixels have large overlaps, which lead to a lot of redundant computation. The proposed algorithms eliminate all the redundant computation in convolution and pooling on images by introducing novel d-regularly sparse kernels. It generates exactly the same results as those by patch-by-patch scanning. Convolution and pooling operations with such kernels are able to continuously access memory and can run efficiently on GPUs. A fraction of patches of interest can be chosen from each training image for backward propagation by applying a mask to the error map at the last CNN layer. Its computation complexity is constant with respect to the number of patches sampled from the image. Experiments have shown that our proposed algorithms speed up commonly used patch-by-patch scanning over 1500 times in both forward and backward propagation. The speedup increases with the sizes of images and patches.

Click to Read Paper and Get Code
Human eyes can recognize person identities based on small salient regions, i.e. human saliency is distinctive and reliable in pedestrian matching across disjoint camera views. However, such valuable information is often hidden when computing similarities of pedestrian images with existing approaches. Inspired by our user study result of human perception on human saliency, we propose a novel perspective for person re-identification based on learning human saliency and matching saliency distribution. The proposed saliency learning and matching framework consists of four steps: (1) To handle misalignment caused by drastic viewpoint change and pose variations, we apply adjacency constrained patch matching to build dense correspondence between image pairs. (2) We propose two alternative methods, i.e. K-Nearest Neighbors and One-class SVM, to estimate a saliency score for each image patch, through which distinctive features stand out without using identity labels in the training procedure. (3) saliency matching is proposed based on patch matching. Matching patches with inconsistent saliency brings penalty, and images of the same identity are recognized by minimizing the saliency matching cost. (4) Furthermore, saliency matching is tightly integrated with patch matching in a unified structural RankSVM learning framework. The effectiveness of our approach is validated on the VIPeR dataset and the CUHK01 dataset. Our approach outperforms the state-of-the-art person re-identification methods on both datasets.

* This manuscript has 14 pages with 25 figures, and a preliminary version was published in ICCV 2013
Click to Read Paper and Get Code
This paper designs a high-performance deep convolutional network (DeepID2+) for face recognition. It is learned with the identification-verification supervisory signal. By increasing the dimension of hidden representations and adding supervision to early convolutional layers, DeepID2+ achieves new state-of-the-art on LFW and YouTube Faces benchmarks. Through empirical studies, we have discovered three properties of its deep neural activations critical for the high performance: sparsity, selectiveness and robustness. (1) It is observed that neural activations are moderately sparse. Moderate sparsity maximizes the discriminative power of the deep net as well as the distance between images. It is surprising that DeepID2+ still can achieve high recognition accuracy even after the neural responses are binarized. (2) Its neurons in higher layers are highly selective to identities and identity-related attributes. We can identify different subsets of neurons which are either constantly excited or inhibited when different identities or attributes are present. Although DeepID2+ is not taught to distinguish attributes during training, it has implicitly learned such high-level concepts. (3) It is much more robust to occlusions, although occlusion patterns are not included in the training set.

Click to Read Paper and Get Code
The key challenge of face recognition is to develop effective feature representations for reducing intra-personal variations while enlarging inter-personal differences. In this paper, we show that it can be well solved with deep learning and using both face identification and verification signals as supervision. The Deep IDentification-verification features (DeepID2) are learned with carefully designed deep convolutional networks. The face identification task increases the inter-personal variations by drawing DeepID2 extracted from different identities apart, while the face verification task reduces the intra-personal variations by pulling DeepID2 extracted from the same identity together, both of which are essential to face recognition. The learned DeepID2 features can be well generalized to new identities unseen in the training data. On the challenging LFW dataset, 99.15% face verification accuracy is achieved. Compared with the best deep learning result on LFW, the error rate has been significantly reduced by 67%.

Click to Read Paper and Get Code
When response variables are nominal and populations are cross-classified with respect to multiple polytomies, questions often arise about the degree of association of the responses with explanatory variables. When populations are known, we introduce a nominal association vector and matrix to evaluate the dependence of a response variable with an explanatory variable. These measures provide detailed evaluations of nominal associations at both local and global levels. We also define a general class of global association measures which embraces the well known association measure by Goodman-Kruskal (1954). The proposed association matrix also gives rise to the expected generalized confusion matrix in classification. The hierarchy of equivalence relations defined by the association vector and matrix are also shown.

Click to Read Paper and Get Code
Statistical features, such as histogram, Bag-of-Words (BoW) and Fisher Vector, were commonly used with hand-crafted features in conventional classification methods, but attract less attention since the popularity of deep learning methods. In this paper, we propose a learnable histogram layer, which learns histogram features within deep neural networks in end-to-end training. Such a layer is able to back-propagate (BP) errors, learn optimal bin centers and bin widths, and be jointly optimized with other layers in deep networks during training. Two vision problems, semantic segmentation and object detection, are explored by integrating the learnable histogram layer into deep networks, which show that the proposed layer could be well generalized to different applications. In-depth investigations are conducted to provide insights on the newly introduced layer.

* refined some typos, ECCV 2016
Click to Read Paper and Get Code
Scene labeling is a challenging classification problem where each input image requires a pixel-level prediction map. Recently, deep-learning-based methods have shown their effectiveness on solving this problem. However, we argue that the large intra-class variation provides ambiguous training information and hinders the deep models' ability to learn more discriminative deep feature representations. Unlike existing methods that mainly utilize semantic context for regularizing or smoothing the prediction map, we design novel supervisions from semantic context for learning better deep feature representations. Two types of semantic context, scene names of images and label map statistics of image patches, are exploited to create label hierarchies between the original classes and newly created subclasses as the learning supervisions. Such subclasses show lower intra-class variation, and help CNN detect more meaningful visual patterns and learn more effective deep features. Novel training strategies and network structure that take advantages of such label hierarchies are introduced. Our proposed method is evaluated extensively on four popular datasets, Stanford Background (8 classes), SIFTFlow (33 classes), Barcelona (170 classes) and LM+Sun datasets (232 classes) with 3 different networks structures, and show state-of-the-art performance. The experiments show that our proposed method makes deep models learn more discriminative feature representations without increasing model size or complexity.

* 13 pages
Click to Read Paper and Get Code
Cascade is a widely used approach that rejects obvious negative samples at early stages for learning better classifier and faster inference. This paper presents chained cascade network (CC-Net). In this CC-Net, the cascaded classifier at a stage is aided by the classification scores in previous stages. Feature chaining is further proposed so that the feature learning for the current cascade stage uses the features in previous stages as the prior information. The chained ConvNet features and classifiers of multiple stages are jointly learned in an end-to-end network. In this way, features and classifiers at latter stages handle more difficult samples with the help of features and classifiers in previous stages. It yields consistent boost in detection performance on benchmarks like PASCAL VOC 2007 and ImageNet. Combined with better region proposal, CC-Net leads to state-of-the-art result of 81.1% mAP on PASCAL VOC 2007.

Click to Read Paper and Get Code