3D MOT에서 JDE 적용이 어려운 이유

Joint Detection and Embedding은 주로 이미지를 대상으로 하는 2D MOT에서 사용된다.

3D MOT에 JDE를 적용하는데 겪는 어려움 및 한계를 알아보자.

 

1.  Poly-MOT: A Polyhedral Framework For 3D Multi-Object Tracking

Due to the data-driven nature of JDT, it is generally less precise and robust than TBD, and consequently, the majority of 3D MOT approaches adhere to the TBD architecture.

 

2. SimpleTrack : Understanding and Rethinking 3D Multi-object Tracking

Many 3D MOT methods are composed of hand-crafted rule-based components. leverage rich features from RGB images for association and life cycle control, specially uses neural networks to handle the feature fusion, association metrics, and tracklet initialization.

 

3D point cloud를 대상으로 하는 3D MOT는 이용할 수 있는 데이터가 motion정보 밖에 없다. 그렇기 때문에 기존의 방식들은 주로 hand-craft module방식(TBD)을 사용하였다. end-to-end로 학습하기 위해서는 더 많은 정보를 담고있는 데이터를 필요로 하였고, 센서퓨전 기법으로 발전하였다. 

 

3. Learnable Online Graph Representations for 3D Multi-Object Tracking

recent work in 2D MOT [13], [14] aims to reduce the amount of heuristics by modeling all tasks in a single learnable pipeline using graph neural networks. However, most of these approaches are limited to the offline setting and driven by appearance-based association that cannot be readily employed in the 3D counterpart.