Core Concepts
The author proposes a fully decoupled end-to-end person search model to optimize performance by separating detection and re-identification tasks. The task-incremental person search network enables independent learning for conflicting objectives, achieving the best results for both sub-tasks.
Abstract
The content discusses the challenges of conflicting objectives in end-to-end person search and introduces a fully decoupled model to address them. By proposing a task-incremental training approach, the model achieves optimal performance for both detection and re-identification tasks. Experimental evaluations demonstrate the effectiveness of the proposed method on datasets like CUHK-SYSU and PRW. The study compares different combinations of detectors and architectures, highlighting the benefits of side-fusion modules in transferring knowledge between tasks. Additionally, efficiency comparisons with previous models show promising results in terms of training time, parameters, and runtime.
Stats
Detection AP: 93.4
Re-ID mAP: 97.6
Training Time: 56.3 hours
Quotes
"The proposed fully decoupled models significantly outperform previous decoupled models on PRW."
"Our proposed method achieves competitive performance without complex model architectures."