search

SysCV / tet

Implementation of Tracking Every Thing in the Wild, ECCV 2022
Apache License 2.0
93 stars 9 forks source link

checkpoint of the model trained LVISv0.5 + COCO dataset #2

Open gaoyuchris opened 1 year ago

gaoyuchris commented 1 year ago

Hi, I wonder can you provide the checkpoint of the model trained on LVISv0.5 + COCO dataset? many thanks.

siyuanliii commented 1 year ago

Thanks. For fintuning on TAO, we freezed the detection parts and only train the track head. Thus, you can download the weights in the repo and drop track head part to get the model that you want.

gaoyuchris commented 1 year ago

Thanks. For fintuning on TAO, we freezed the detection parts and only train the track head. Thus, you can download the weights in the repo and drop track head part to get the model that you want.

Thanks for your reply @siyliepfl . I understand that the track head will also be trained in the training on LVISv0.5 + COCO dataset, which take the key frame itself as its ref frame (according to https://github.com/SysCV/tet/blob/2c18172607d6fc526083c67d93e80104fdc98577/configs/tao/cem_r101_lvis.py#L31). Thus the weights in the repo with the track head part dropped are not the weights I want.