M3L: Multi-modality mining for metric learning in person re-Identification

Highlights：

• A multi-modality mining approach is proposed to automatically discov- er multiple modalities of illumination changes in long-run surveillance scenes for person re-identification.

• A shift-invariant property of the RGB values from related views in log-chromacticity space is discovered.

• The proposed multi-modality mining algorithm can achieve boosted performance against conventional global metric learning algorithms.

• In scenarios with more complicated illumination changes, our method can achieve more distinct improvements.

• The proposed multi-modality mining approach is independent of specif- ic metric learning algorithms, so any advanced metric learning method can exploit our framework to obtain superior performance in future.

摘要

•A multi-modality mining approach is proposed to automatically discov- er multiple modalities of illumination changes in long-run surveillance scenes for person re-identification.•A shift-invariant property of the RGB values from related views in log-chromacticity space is discovered.•The proposed multi-modality mining algorithm can achieve boosted performance against conventional global metric learning algorithms.•In scenarios with more complicated illumination changes, our method can achieve more distinct improvements.•The proposed multi-modality mining approach is independent of specif- ic metric learning algorithms, so any advanced metric learning method can exploit our framework to obtain superior performance in future.

论文评审过程：Received 27 December 2016, Revised 12 August 2017, Accepted 27 September 2017, Available online 6 October 2017, Version of Record 8 January 2018.