Video-Based Person Re-Identification via Self Paced Weighting

Wenjun Huang, Chao Liang, Yi Yu, Zheng Wang, Weijian Ruan, Ruimin HuAAAI 2018 · DOI

Abstract

Video-based person re-identification (re-id) faces significant challenges due to unregulated sequences containing occlusions and noise. This paper presents a Self Paced Weighting (SPW) framework designed to automatically identify and down-weight these "noisy" segments. By decomposing videos into stable fragments and applying a curriculum-learning-inspired outlier detection mechanism, the proposed system prioritizes high-quality visual data, achieving state-of-the-art accuracy on benchmark datasets.

Introduction

The primary goal of person re-identification is to match individuals across non-overlapping camera views. While video sequences offer rich temporal and multi-view information compared to static images, they are often corrupted by real-world factors such as occlusions, motion blur, and background clutter. Traditional algorithms frequently treat all frames with equal importance, allowing low-quality data to degrade the final feature representation. This often results in a reduced margin between inter-class and intra-class distances, complicating the matching process.

Figure 1: Comparison of intra- and inter-class distances between original sequences and cleaned sub-sequences.

Proposed Methodology

The SPW framework addresses sequence noise through a three-stage pipeline:

1. Video Sequence Segmentation

Rather than processing a video as a monolithic block, the authors employ a Sequence Stability Measure (SSM). This measure analyzes occlusion patterns between consecutive frames to identify "stationary points"—moments of relative visual stability. The video is then segmented into sub-sequences centered on these points, effectively isolating "clean" fragments from "noisy" ones.

Figure 3: The sequence segmentation process using the Sequence Stability Measure (SSM) to identify stable fragments.

2. Self Paced Outlier Detection (SPOD)

Inspired by curriculum learning, SPOD quantifies the "noise degree" of each sub-sequence. It utilizes a ranking loss function to identify outliers. Sub-sequences exhibiting a loss higher than a dynamically learned threshold are flagged as noisy, allowing the model to focus on reliable data during the learning process.

3. Weighted Multi-pair Distance Metric

During the matching phase between probe and gallery videos, the system applies the weights derived from the SPOD stage. This ensures that high-quality, clean fragments contribute significantly to the final similarity score, while the influence of noisy fragments is minimized.

Experimental Results

The SPW method was evaluated on two major benchmarks: PRID 2011 and iLIDS-VID.

  • iLIDS-VID: In this challenging airport environment characterized by heavy occlusions, SPW achieved a Rank-1 accuracy of 69.3%, significantly outperforming previous state-of-the-art (SOTA) methods like ASTPN (62.0%) and TDNN (58.0%).
  • PRID 2011: On this cleaner dataset, the method reached 83.5% Rank-1 accuracy, surpassing the previous best of 79.4%.

The framework demonstrated robust "plug-and-play" capabilities, improving performance across various feature extractors, including HOG3D and deep IDE features.

Conclusion

The Self Paced Weighting framework highlights that in video-based re-id, data quality is more critical than data quantity. By intelligently segmenting sequences and filtering outliers through a curriculum-based approach, the system effectively mitigates the impact of occlusions. Future work may explore the integration of more advanced spatial-temporal descriptors and the adaptation of this weighting logic to unsupervised learning environments.

Summary updated

BibTeX
@inproceedings{huang2018video,
  author = {Wenjun Huang and Chao Liang and Yi Yu and Zheng Wang and Weijian Ruan and Ruimin Hu},
  title = {{Video-Based Person Re-Identification via Self Paced Weighting}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.11857},
}
Figure