MCG-NJU/MOTIP

What it solves

MOTIP addresses the challenge of Multiple Object Tracking (MOT) by reframing it as an in-context ID prediction problem. Instead of traditional tracking-by-detection methods, it aims to consistently assign ID labels to detections across frames to maintain stable object identities over time.

How it works

MOTIP treats tracking as a sequence-to-sequence problem. Given a set of trajectories containing ID information from previous frames, the model directly decodes the ID labels for detections in the current frame. This approach simplifies the tracking process by treating the ID assignment as a prediction based on the context of existing trajectories.

Who it’s for

Computer vision researchers and developers working on multiple object tracking, specifically those looking for an alternative to traditional association-based tracking pipelines.

Highlights

• In-context ID prediction: Reframes MOT as a prediction task rather than a simple association task.
• Nearly real-time tracking: Supports fast inference, including FP16 precision for increased speed.
• DETR-based architecture: Built upon Deformable DETR and MOTR.
• CVPR 2025: The method is presented in a CVPR 2025 paper.