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
3.55k stars 594 forks source link

why , when using mmdetection pretrained weight for detection is not giving good results in tracking inference ? #563

Open Balakumaran-kandula opened 2 years ago

Balakumaran-kandula commented 2 years ago

I have tried with the different pretrained weights given from mmdetection models ( e.g) (https://download.openmmlab.com/mmdetection/v2.0/cascade_rcnn/cascade_mask_rcnn_x101_64x4d_fpn_mstrain_3x_coco/cascade_mask_rcnn_x101_64x4d_fpn_mstrain_3x_coco_20210719_210311-d3e64ba0.pth) combined with reid model weights (e.g) (https://download.openmmlab.com/mmtracking/mot/reid/reid_r50_6e_mot17-4bf6b63d.pth ) for multi object tracking inference , but the output was random bounding boxes , does it have to be pretrained on tracking datasets and models to work fine with mmtracking inference ?

JingweiZhang12 commented 2 years ago

Fintine on tracking dataset will bring more performance gain, but the original model from mmdetection is not so bad to output random bounding boxes for all objects. Maybe the keys between the checkpoint from mmdetection and the model in mmtracking are different. The model in mmdetection need to be converted to be modle.detector in mmtracking. Please check its config, https://github.com/open-mmlab/mmtracking/blob/master/docs/en/tutorials/config_mot.md

Balakumaran-kandula commented 2 years ago

mode.detector ? ( there is no such command) . I guess you are saying about model.detector , i have just gave the pretrained weight link in the config of mot ( tracktor) model ,it is loading the detector perfectly . but the output is not good ,