Journal / Conference
IEEE International Conference on Computer Vision (ICCV2019)
[PDF link: here]
[Code link: here]
Keywords
Person Re-ID, Spectral Feature Transformation, Post-processing
Abstract
With the surge of deep learning techniques, the field of person re-identification has witnessed rapid progress in recent years. Deep learning based methods focus on learning a discriminative feature space where data points are clustered compactly according to their corresponding identities. Most existing methods process data points individually or only involves a fraction of samples while building a similarity structure. They ignore dense informative connections among samples more or less. The lack of holistic observation eventually leads to inferior performance. To relieve the issue, we propose to formulate the whole data batch as a similarity graph. Inspired by spectral clustering, a novel module termed Spectral Feature Transformation is developed to facilitate the optimization of group-wise similarities. It adds no burden to the inference and can be applied to various scenarios. As a natural extension, we further derive a lightweight re-ranking method named Local Blurring Re-ranking which makes the underlying clustering structure around the probe set more compact. Empirical studies on four public benchmarks show the superiority of the proposed method.
Method/Framework
We adopt the output of the final global average pooling layer as image embedding for retrieval. Spectral feature transformation is performed on the embeddings of the data batch. Subsequently, a classifier is imposed on the transformed feature. Suppose is the final embedding of a training batch. We also combine the model with an extra classification branch. Parameters are shared between the two classifiers.
As for post-processing stage, given a probe image, images in the gallery are ranked according to the cosine similarity with it. Then, we collect features of top-n entries and perform spectral feature transformation on them. Finally, the top-n rank list is recomputed based on the similarity derived from transformed features.
Experiments
To validate the effectiveness of the proposed method, we conduct extensive experiments on four popular person re-identification benchmarks, i.e, Market-1501, DukeMTMC-reID, CUHK03 and MSMT17.
Highlight
- To efficiently capture more informative structure, we form the data in one batch into a similarity graph. Inspired by spectral clustering, a novel feature transformation is proposed which facilitates the optimization of group-wise similarities on the graph. It introduces no extra cost to the inference and can be readily adapted to other tasks which requires embedding.
- A lightweight re-ranking method is naturally derived. It makes the underlying clustering structure more compact in the neighborhood of the probe set.
- Extensive experiments validate the effectiveness of our method. Competitive performances are achieved on all four public benchmarks.
Citation
@inproceedings{luo2019spectral,
title={Spectral feature transformation for person re-identification},
author={Luo, Chuanchen and Chen, Yuntao and Wang, Naiyan and Zhang, Zhaoxiang},
booktitle={Proceedings of the IEEE International Conference on Computer Vision},
year={2019}
}
