search

open-mmlab / mmtracking

OpenMMLab Video Perception Toolbox. It supports Video Object Detection (VID), Multiple Object Tracking (MOT), Single Object Tracking (SOT), Video Instance Segmentation (VIS) with a unified framework.
https://mmtracking.readthedocs.io/en/latest/
Apache License 2.0
3.58k stars 598 forks source link

reid accuracy #531

Open CarlHuangNuc opened 2 years ago

CarlHuangNuc commented 2 years ago

Why retrain reid config alway get very low accuracy(mAP:50),

From your log & pretrain model, it can achieve about 96% mAP.

dyhBUPT commented 2 years ago

Hi, we can't understand your question accurately. We recommend using the issue template.

Best wishes.

biyoml commented 2 years ago

I had the same issue. Did you solve it?

dyhBUPT commented 2 years ago

Can your retrained ReID model achieve the same MOT metrics for Tracktor? I mean, you just replace the reid ckpt with your retrained results and leave detector and ohter hyper-parameters unchanged.

biyoml commented 2 years ago

Although my retrained ReID checkpoint only had ~60% mAP during training, the Tracktor (tracktor_faster-rcnn_r50_fpn_4e_mot17-private-half) using my checkpoint can achieve 63.7% MOTA on MOT17.

JinZhangYu commented 2 years ago

same problem here, any fix now?

JinZhangYu commented 2 years ago

https://download.openmmlab.com/mmtracking/fp16/reid_r50_fp16_8x32_6e_mot17_20210731_033055.log.json This log shows each epoch has 8000 iterations under GPU8, but master tools/dataset_converters/mot/mot2reid.py cannot generate so many crop images under MOT17.

I use 4 GPUs, and each epoch only has 873 iterations.