End-to-end Multi-object Detection and Re-identification with Similar Appearance

In the field of computer vision, detecting and re-identifying multiple objects with similar appearances is an important and challenging task. Traditional multi-object tracking methods usually adopt a serial detection and re-identification paradigm. Although this design offers flexibility to accommodate various detectors, the tracking speed is constrained by the relatively slow re-identification model, preventing it from fully leveraging the fast inference capabilities of the detector. To better balance speed and accuracy, end-to-end detection and re-identification models have garnered significant attention in the field of multi-object tracking. For scenes with similar appearance, we propose an improved end-to-end detection and re-identification model based on an encoder-decoder structure, in which the encoder part introduces a triple attention mechanism to capture dependencies in multiple dimensions and extract features with rich representation capabilities; the decoder part adopts a decoupled multi-branch structure and introduces deformable convolution in the re-identification branch to dynamically sample feature points. Experiments indicate that the improved method boosts the re-identification IDF1 by 6.9%, with detection performance remaining nearly unaffected. The improved method effectively improves the association performance under similar appearance and can be applied to multi-object tracking in highly dynamic and complex scenes.