MODEL_ZOO.md

MODEL ZOO

• The experiments are run with PyTorch 1.0, CUDA 10.0, and CUDNN 7.5.
• Training and testing times are measured on our servers with V100s GPUs.
• The models can be downloaded directly from Google drive.

KITTI MOTS

Model	GPUs	Train time	Test time	Validation HOTA	Test HOTA	Download
kitti_mots_sch_fulltrain	2	-	0.19s	-	53.1/67.1	model
kitti_mots_sch_half	1	-	0.19s	63.5/78.5	-	model

Notes

• *_half corresponds to the half-half video train/ val split mentioned in the paper.
• *_fulltrain corresponds to train on the full training set.
• HOTA are report in pedestrian/car.
• all models are fine tune on ctdet_coco_dla_2x.pth provided by CenterNet.

MOT17

Model	GPUs	Train time	Test time	Validation HOTA	Test HOTA	Download
mot17_fulltrain_sch_ch	2	-	-	-	53.4	model
mot17_half_sch	2	-	-	-	-	model
crowdhuman_sch	2	-	-	-	-	model

Notes

• *_half corresponds to the half-half video train/ val split mentioned in the paper.
• *_fulltrain corresponds to train on the full training set and submit to leaderboard using public detections.
• mot17_half/ mot17_fulltrain are finetuned on the crowdhuman model, and mot17_half_sc/ mot17_fulltrain_sc are trained from ImageNet initialization.
• The crowdhuman model is trained on CrowdHuman dataset with similar setting as CenterTrack.