Introduction

主要用来系统解释tracking流程。

contributions are as follow:

1. pipeline of “tracking-by-detection” ⇒ 4 components
   1. analyze the connections between each component and failure cases.
2. propose corresponding treatments for each module ⇒ a strong yet simple baseline.
3. analyze existing 3D MOT benchmarks and explain the potential influences of their designs.

1. Pre-processing of Input Detections
   1. kill bounding boxes with low quality or severe overlapping
   2. use CenterPoint detection
2. Motion Model
   1. in Prediction & motion model update
   2. Kalman filter or velocities
3. Association
   1. associates the detections with tracklets
   2. ⇒ Generalized IoU metric
4. Life Cycle Management
   1. controls the “birth”, “death” and “output” policies.
   2. ⇒ Two-stage Association
   3. the same as same as AB3DMOT [37]: 3 hits to start outputting a tracker and consecutive 2 misses terminates a tracklet

Pipeline: 4

Pre-processing

1. Score vs NMS ⇒ NMS
   1. inaccurate input detections ⇒ contaminate the association.
   2. high score threshold ⇒ sacrifice the recall.

Tab 1 & 2Ablation for NMS on nuScenes & WOD.

Motion Model

1. Kalman filter vs Constant velocity ⇒ what? it depends on the datasets.

   1. similar performance in general

   2. KF fits better for high frequency cases because of more predictable motions

      Tab 3: KF, CV, KF PD ⇒ refinement is important for

   3. CV is more robust for low-frequency scenarios with explicit speed prediction

Association