Guanwen Zhang

Adaptive Metric Learning in Local Distance Comparison for People Re-identification

A local distance comparison for multiple-shot people re-identification based on a new adaptive metric learning method is introduced in this paper. There exist two intrinsic issues in multiple-shot person re-identification: Large variances in view point, illumination, and non-rigid deformation are included in the image set of the same person, only a few training data for learning tasks are available in a realistic re-identification scenario. We deal with the multimodal property of peoples appearance distribution caused by the first issue by using a local distance comparison approach. Since the capability of the local distance comparison highly depends on the choice of distance metric, we also introduce an adaptive learning method to learn an appropriate distance metric and use it to find and compute local neighbors effectively. This adaptive learning method is able to solve the over fitting problem caused by the second issue, through leveraging the generic knowledge of re-identification together with the specific information of the target task. We evaluated our approach on public benchmark datasets, and confirmed its superiority as compared to conventional approaches.

Solving Occlusion Problem in Pedestrian Detection by Constructing Discriminative Part Layers

Occlusion handling is one of the most challenging issues for pedestrian detection, and no satisfactory achievement has been found in this issue yet. Using human body parts has been considered as a reasonable way to overcome such an issue. In this paper, we propose a brand new approach based on the fusion of Mid-level body part mining and Convolutional Neural Network (CNN) to solve this problem, named DP-CNN(Discriminative Parts CNN). Two main discussions are included in this paper. First, we take an exhaustive analysis on how to mine useful body parts that contribute to pedestrian detection. Multiple ingredients (e.g. feature representation, pedestrian attributes) are analyzed through a wide range of experiments. Second, we convert the part detectors to the middle layer of CNN and re-train the model to get a better adaption of the dataset. Compare to existing approaches based on fine-tuning CNN models, our method is not only robust to occlusion handling but also has a smaller computational cost.

How to initialize the CNN for small datasets: Extracting discriminative filters from pre-trained model

In this paper, we study how to initialize the convolutional neural network (CNN) model for training on a small dataset. Specially, we try to extract discriminative filters from the pre-trained model for a target task. On the basis of relative entropy and linear reconstruction, two methods, Minimum Entropy Loss (MEL) and Minimum Reconstruction Error (MRE), are proposed. The CNN models initialized by the proposed MEL and MRE methods are able to converge fast and achieve better accuracy. We evaluate MEL and MRE on the CIFAR10, CIFAR100, SVHN, and STL-10 public datasets. The consistent performances demonstrate the advantages of the proposed methods.

Add-On Strategies for Fine-Grained Pedestrian Classification

In this paper, we present four add-on strategies for the fine-grained pedestrian classification task. These strategies are: (1) super-resolution based image preprocessing, which helps to recover the image details; (2) patch dividing based deep feature extraction, which extracts features in a way that preserves the spatial layout of input images; (3) pose- wise classifier sharing, which learns robust classifiers and makes robust predictions using pose information; and (4) graphical model based inference, which utilizes the interdependence between different subcategories to update raw estimations. The proposed strategies are independent and flexible, which make it easy to implement them in practice. We evaluated these strategies on the CRP dataset and confirmed that all of them lead to improvements over the baseline. We also confirmed an improvement over the state-of-the-art when all strategies are combined together.

People re-identification using two-stage transfer metric learning

With assumptions that people usually do not change their clothes during an observation period, people appearance data are easily outdated in re-identification applications. This raises the over-fitting problem because only a few training data are available for learning statistical models. In this paper, we propose a two-stage transfer metric learning approach for multiple-shot people re-identification to tackle this small training data problem. In the first stage, we transfer the generic knowledge from a large existing dataset, and in the second stage, we transfer the learned distance metric for each probe-specific person using the side-information. Experimental results on several public benchmark datasets show that our proposed approach is superior over conventional approaches.