Skip to content

Latest commit

 

History

History
43 lines (28 loc) · 4.71 KB

MODEL_ZOO.md

File metadata and controls

43 lines (28 loc) · 4.71 KB

Model Zoo

MOTIP

DanceTrack

Method Extra Data Traj Aug Resources HOTA DetA AssA
MOTIP no yes pre-train | config | checkpoint 69.6 80.4 60.4
MOTIP no no pre-train | config | checkpoint 65.2 80.4 53.1

SportsMOT

Method Extra Data Traj Aug Resources HOTA DetA AssA
MOTIP no yes pre-train | config | checkpoint 72.6 83.5 63.2
MOTIP no no pre-train | config | checkpoint 70.9 83.7 60.1

BFT

Method Extra Data Traj Aug Resources HOTA DetA AssA
MOTIP no yes pre-train | config | checkpoint 70.5 69.6 71.8
MOTIP no no pre-train | config | checkpoint 71.3 69.2 73.7

NOTE:

  1. Traj Aug is an abbreviation for Trajectory Augmentation in the paper.
  2. You could also load previous checkpoints for inference from prev-engine branch, using runtime parameter --use-previous-checkpoint True. You may need to pass additional parameters to bridge the difference in the experimental setups. Typically, --rel-pe-length and --miss-tolerance.

DETR

You can directly download the pre-trained DETR weights used in our experiment here (recommended). Or you can choose to follow the guidance to perform pre-training yourself.

Model Name Target Dataset Extra Data Resources
r50_deformable_detr_coco COCO no checkpoint
r50_deformable_detr_coco_dancetrack DanceTrack no config | checkpoint
r50_deformable_detr_coco_sportsmot SportsMOT no config | checkpoint
r50_deformable_detr_coco_bft BFT no config | checkpoint