Zemin Wu

Efficient person re-identification by hybrid spatiogram and covariance descriptor

Feature and metric researchings are two vital aspects in person re-identification. Metric learning seems to have gained extra advantage over feature in recent evaluations. In this paper, we explore the neglected potential of feature designing for re-identification. We propose a novel and efficient person descriptor, which is motivated by traditional spatiogram and covariance descriptors. The spatiogram feature accumulates multiple spatial histograms of different image regions from several color channels and then extracts three descriptive sub-features. The covariance feature exploits several colorspaces and intensity gradients as pixel features and then extracts multiple statistical feature vectors from a pyramid of covariance matrices. Moreover, we also propose an effective and efficient multi-shot re-id metric without learning, which fuses the residual and coding coefficients after collaboratively coding samples on all person classes. The proposed descriptor and metric are evaluated with current methods on benchmark datasets. Our methods not only achieve state-of-the-art results but also are straightforward and computationally efficient, facilitating real-time surveillance applications such as pedestrian tracking and robotic perception in various dynamic scenes.

Person Re-Identification Based on Spatiogram Descriptor and Collaborative Representation

Feature and metric designing are two vital aspects in person re-identification. In this letter, we firstly propose a novel spatiogram based person descriptor. Such spatiograms of different image regions from several color channels are calculated and accumulated to create a histogram vector and two distinctive spatial statistical vectors. Secondly, through further investigating the multi-shot set-based metric based on the recent collaborative representation model, we propose an effective and efficient multi-shot metric, which fuses the residual and coding coefficients after collaboratively coding samples on all person classes. Finally, we evaluate the proposed descriptor and metric with other published methods on benchmark datasets. Our methods not only achieve state-of-the-art results but also are novel, straightforward and computationally efficient, which will facilitate the real-time surveillance applications such as pedestrian tracking.

Person Re-identification with Hierarchical Deep Learning Feature and efficient XQDA Metric

Feature learning and metric learning are two important components in person re-identification (re-id). In this paper, we utilize both aspects to refresh the current State-Of-The-Arts (SOTA). Our solution is based on a classification network with label smoothing regularization (LSR) and multi-branch tree structure. The insight is that some middle network layers are found surprisingly better than the last layers on the re-id task. A Hierarchical Deep Learning Feature (HDLF) is thus proposed by combining such useful middle layers. To learn the best metric for the high-dimensional HDLF, an efficient eXQDA metric is proposed to deal with the large-scale big-data scenarios. The proposed HDLF and eXQDA are evaluated with current SOTA methods on five benchmark datasets. Our methods achieve very high re-id results, which are far beyond state-of-the-art solutions. For example, our approach reaches 81.6%, 96.1% and 95.6% Rank-1 accuracies on the ILIDS-VID, PRID2011 and Market-1501 datasets. Besides, the code and related materials (lists of over 1800 re-id papers and 170 top conference re-id papers) are released for research purposes.

Video saliency from compressed domain coding length

In order to simulate this feature and detect the salient region rapidly, we propose the Spatial-Temporal Feature in Compress Domain (STFCD) model. By respectively using H.264 residual coding length and motion vector coding length, we simulate the salient stimulus intensity and then get video saliency features. Finally, we use the linear weighted fusion algorithm to get the final video saliency maps. Experimental results on three public datasets demonstrate that our model outperforms state-of-the-art methods.

Salient object detection based on discriminative boundary and multiple cues integration

Abstract. In recent years, many saliency models have achieved good performance by taking the image boundary as the background prior. However, if all boundaries of an image are equally and artificially selected as background, misjudgment may happen when the object touches the boundary. We propose an algorithm called weighted contrast optimization based on discriminative boundary (wCODB). First, a background estimation model is reliably constructed through discriminating each boundary via Hausdorff distance. Second, the background-only weighted contrast is improved by fore-background weighted contrast, which is optimized through weight-adjustable optimization framework. Then to objectively estimate the quality of a saliency map, a simple but effective metric called spatial distribution of saliency map and mean saliency in covered window ratio (MSR) is designed. Finally, in order to further promote the detection result using MSR as the weight, we propose a saliency fusion framework to integrate three other cues—uniqueness, distribution, and coherence from three representative methods into our wCODB model. Extensive experiments on six public datasets demonstrate that our wCODB performs favorably against most of the methods based on boundary, and the integrated result outperforms all state-of-the-art methods.

MSR: A Simple and Effective Metric for Visual Saliency Map Fusion

Various kinds of models have been proposed for saliency detection, but each has its limitations in application. Fusing multiple complementary models is expected to improve the performance. However, most of the fusion methods treat individual model equally and are greatly degraded by poor saliency map. In this paper, we firstly present a no-reference metric to assess the quality of saliency map. Then, a fusion framework is constructed by weighted averaging good saliency maps and filtering out poor ones. In this framework, candidate models can be selected with preference to fast ones. Experimental results on two public datasets show that our method not only outperforms state-of-art unsupervised saliency model but also is more robust than present fusion algorithms.

More successful recognition: Seeking the relation of video object detection performance with video coding parameters

Video object recognition is an importance topic in both defense and civilian applications. However, object recognition guarantee is not considered in current video coding standard H.264. For this issue, experiments based on H.264/AVC codec are designed to measure the affection of coding parameters on recognition performance. The quantitative functions of recognition performance with coding rate and resolution ratio are given. More practically, such functions can be used for optimal tradeoff between the channel bandwidth and video object recognition performance when efficient video data transfer is needed.

Person re-identification by multi-channel feature cascading

Person re-identification is a non-trivial problem due to many challenging factors such as different views or varying illumination. The person re-identification problem is discussed in views of feature cascading and combining in this paper. After a detailed analysis of the disadvantages of the ELF (Ensemble of Localized Features) method, features by cascading multiple channels at proper regions are investigated on the reidentification performance. In the proposed exploring framework, its found that the cascading histogram feature with YCbCrHS channels within interior regions can achieve excellent performance. The proposed excellent multi-channel cascading feature (MCCF) is then combined with other features for further improvement. Extensive validation and comparative experiments were conducted on the public gallery VIPeR. And the experimental results show that the proposed feature MCCF and the combined feature can achieve comparable or better performance on person re-identification than other state-of-the- art features.