Zhun Zhong

Re-ranking Person Re-identification with k-Reciprocal Encoding

When considering person re-identification (re-ID) as a retrieval process, re-ranking is a critical step to improve its accuracy. Yet in the re-ID community, limited effort has been devoted to re-ranking, especially those fully automatic, unsupervised solutions. In this paper, we propose a k-reciprocal encoding method to re-rank the re-ID results. Our hypothesis is that if a gallery image is similar to the probe in the k-reciprocal nearest neighbors, it is more likely to be a true match. Specifically, given an image, a k-reciprocal feature is calculated by encoding its k-reciprocal nearest neighbors into a single vector, which is used for re-ranking under the Jaccard distance. The final distance is computed as the combination of the original distance and the Jaccard distance. Our re-ranking method does not require any human interaction or any labeled data, so it is applicable to large-scale datasets. Experiments on the large-scale Market-1501, CUHK03, MARS, and PRW datasets confirm the effectiveness of our method.

Class-specific object proposals re-ranking for object detection in automatic driving

Object proposal generation is an important step in object detection, obtaining high-quality proposals can effectively improve the performance of detection. In this paper, we propose a semantic, class-specific approach to re-rank object proposals, which can consistently improve the recall performance even with fewer proposals. Specifically, we first extract features for each proposal including semantic segmentation, stereo information, contextual information, CNN-based objectness and low-level cue, and then score them using class-specific weights learned by Structured SVM. The advantages of the proposed model are two-fold: 1) it can be easily merged to existing generators with few computational costs, and 2) it can achieve high recall rate under strict critical even using fewer proposals. Experimental evaluation on the KITTI benchmark demonstrates that our approach significantly improves existing popular generators on recall performance. Moreover, in the experiment conducted for object detection, even with 1500 proposals, our approach can still have higher average precision (AP) than baselines with 5000 proposals.

Detecting ground control points via convolutional neural network for stereo matching

In this paper, we present a novel approach to detect ground control points (GCPs) for stereo matching problem. First of all, we train a convolutional neural network (CNN) on a large stereo set, and compute the matching confidence of each pixel by using the trained CNN model. Secondly, we present a ground control points selection scheme according to the maximum matching confidence of each pixel. Finally, the selected GCPs are used to refine the matching costs, then we apply the new matching costs to perform optimization with semi-global matching algorithm for improving the final disparity maps. We evaluate our approach on the KITTI 2012 stereo benchmark dataset. Our experiments show that the proposed approach significantly improves the accuracy of disparity maps.

Attention guided U-Net for accurate iris segmentation

Abstract Iris segmentation is a critical step for improving the accuracy of iris recognition, as well as for medical concerns. Existing methods generally use whole eye images as input for network learning, which do not consider the geometric constrain that iris only occur in a specific area in the eye. As a result, such methods can be easily affected by irrelevant noisy pixels outside iris region. In order to address this problem, we propose the ATTention U-Net (ATT-UNet) which guides the model to learn more discriminative features for separating the iris and non-iris pixels. The ATT-UNet firstly regress a bounding box of the potential iris region and generated an attention mask. Then, the mask is used as a weighted function to merge with discriminative feature maps in the model, making segmentation model pay more attention to iris region. We implement our approach on UBIRIS.v2 and CASIA.IrisV4-distance, and achieve mean error rates of 0.76% and 0.38%, respectively. Experimental results show that our method achieves consistent improvement in both visible wavelength and near-infrared iris images with challenging scenery, and surpass other representative iris segmentation approaches.

Multi-matched Similarity: A New Method for Image Retrieval

In Bag-of-Words BoW based image retrieval, soft assignment SA assigns R-nearest visual words to a feature, which significantly enhances the performance of image retrieval. However, it requires to calculate the weight of each visual word according to the distance from feature to visual word. This method sometimes loses its power when the codebook size is small, since a smaller codebook will cause larger quantization error and lead the distance to be more imprecise. In this paper, instead of depending on distance, we present a novel method to calculate the similarity between features by counting the number of identical visual words assigned to them. We describe how to create the inverted index, and weight the score between matched features. We evaluate the proposed Multi-Matched Similarity MMS method on Holidays and Ukbench datasets. Experimental results demonstrate that our method significantly improves the retrieval performance and outperforms the SA approach on both datasets.