Computer Vision Lab

Re-identification Based on Spatial-temporal Fusion Network

Hye-Geun Kim¹, You-Kyoung Na¹, Hae-Won Jo¹, Yong-Hyuk Moon² and Yeong-Jun Cho¹

¹Chonnam National University, ²Sungshin Women's University

International Conference on Pattern Recognition (ICPR) 2024, ⭐Oral presentation
Computer Vision and Image Understanding, in revision, 2025

[Paper-journal][Paper-conference][arxiv][Code]

ABSTRACT

Vehicle re-identification (ReID) in a large-scale camera network is important in public safety, traffic control, and security. However, due to the appearance ambiguities of vehicle, the previous appearance-based ReID methods often fail to track vehicle across multiple cameras. To overcome the challenge, we propose a spatial-temporal vehicle ReID framework that estimates reliable camera network topology based on the adaptive Parzen window method and optimally combines the appearance and spatial-temporal similarities through the fusion network. Based on the proposed methods, we performed superior performance on the public dataset (VeRi776) by 99.64% of rank-1 accuracy. The experimental results support that utilizing spatial and temporal information for ReID can leverage the accuracy of appearance-based methods and effectively deal with appearance ambiguities.

OVERVIEW

We observed that alleviating appearance ambiguity is difficult using only the appearance-based ReID model. Many studies have used spatial-temporal information to alleviate appearance ambiguity; however, appearance and spatial-temporal information are not optimized. Therefore, we propose ReID based on the proposed spatial-temporal fusion network (FusionNet) to overcome the limitations. First, we estimated the camera network topology using the proposed adaptive Parzen window that is robust to outliers and sparse responses between camera pairs.

Figure 1. The overall framework for re-identification with spatial-temporal information.

While the Parzen window method efficiently estimates continuous PDFs from discrete histograms, employing a single kernel across diverse histograms from various camera pairs is unreasonable. The strength of the spatial-temporal connection between cameras can be determined by how many objects pass through those cameras during a certain period. For example, few positive pairs between two cameras indicate weak connectivity. Nevertheless, the Parzen window method extremely enlarges those small responses with a small σ value, as depicted in the orange line in Fig. 2 (c). In that case, it is better to use a large σ value to avoid overfitting the distribution for noise and outliers.

To overcome the limitation of the original Parzen window method, we newly propose an adaptive Parzen window by setting various σij values for the camera pairs (ci,cj). To this end, we designed an adaptive standard deviation according to the different strengths of the camera connectivity as shown in Fig. 2 blue lines.

Figure 2. Examples of estimated transition time distributions between camera pairs.

RESULTS

Table 1. Performance comparisons on the VeRi776 and Market1501 datasets. †and ∗ indicate the spatial-temporal approach and re-ranking, respectively. The best and second best performances are marked in bold and underline.

Figure 3. Qualitative ReID results of the baseline and proposed methods (Ours). The baseline model is an appearance-based method. Green and red boxes denote true and false matching. Compared to the baseline method, the proposed method can match true positive pairs despite similar appearances and overcome the appearance ambiguities due to the spatial-temporal information (best viewed in color).

🖇️ Citation

@inproceedings{kim2025re, title={Re-Identification Based on the Spatial-Temporal Fusion Network},
author={Kim, Hye-Geun and Na, You-Kyoung and Joe, Hae-Won and Moon, Yong-Hyuk and Cho, Yeong-Jun},
booktitle={International Conference on Pattern Recognition},
pages={421--436},
year={2025},
organization={Springer}
}

Page updated

Google Sites

Report abuse