Project / 02 · 2026
corrMOTRResearch prototype
Explicit State Disentanglement for End-to-End Multi-Object Tracking
A MOTR-oriented research system that separates geometry/motion, appearance identity, and existence/lifecycle into explicit state domains.
Research question
Can persistent identity state be separated from transient observation state without losing the simplicity of end-to-end set prediction?
This page treats the project as a system and a research record, not as a list of technologies.
System
Method & contribution
Organized the research code around explicit state tuples, TrackPool history, motion and appearance priors, dual Det/Trk decoders, evidence fusion, and separated DET/TRACK/FUSION supervision.
- 组织显式 state tuple 与 TrackPool history
- 实现运动、外观先验及 Det / Trk 双解码器实验分支
- 拆分 DET / TRACK / FUSION 监督并记录不稳定设置
- Expose state domains instead of hiding them inside one query
- Use uncertainty and evidence to regulate write-back
- Keep experimental branches and failure notes visible
- Separate research hypotheses from validated claims
Evidence
- Research-oriented implementation with explicit state domains
- Architecture, training, and known-limitations records
- Research-oriented implementation
- Architecture and training documentation
- Webcam and evaluation entry points
- Explicit known-limitations record
Some branches are research hypotheses and combined loss settings can be unstable; the repository records these states instead of presenting every path as fully validated.