A Closer Look at the Joint Training of Object Detection and Re-Identification in Multi-Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Multi-object tracking (MOT) benefits from unified object detection and re-identification (ReID) networks for efficiency.
• Joint training in this multi-task setting presents significant challenges, often leading to suboptimal performance due to inherited training practices from object detection.

Purpose of the Study:

• To investigate and improve the joint training of detection and ReID for MOT by focusing on label assignment and loss functions.
• To develop tailored methods that address the specific needs of MOT training, moving beyond standard object detection techniques.

Main Methods:

• Proposed an Identity-aware Label Assignment strategy that incorporates ReID costs into the assignment process, ensuring unambiguous positive samples.
• Introduced a Discriminative Focal Loss that leverages ReID predictions to focus training on discriminative samples.
• Integrated these novel techniques into the FairMOT baseline architecture.

Main Results:

• Achieved substantial improvements on MOT16/17/20 benchmarks, with up to 7.0 MOTA and 54.1% IDs gains.
• Demonstrated superior performance compared to methods using inherited object detection training practices.
• Maintained favorable inference speeds while enhancing tracking accuracy.

Conclusions:

• The proposed Identity-aware Label Assignment and Discriminative Focal Loss are effective for joint detection and ReID in MOT.
• Tailored training strategies are crucial for optimizing multi-task learning in MOT, outperforming generic approaches.
• These advancements offer a more robust and efficient solution for multi-object tracking.