Open qihao-huang opened 3 years ago
ok, I got it, you only take "person" into consideration. Perhaps, a complete adoption with previous MaskTrackRCNN
will be better.
https://github.com/Exely/masktrackrcnn-mmdet2.0/blob/83fcc2b0342f17e79d17d151ab8b1b11dc4359c8/mmdet/datasets/ytvos.py#L21
Thanks for following up! Exactly, the code is adapted for the tianchi competition. We only consider segementing humans in our code by setting one person class. You can simply modify the parameter to adapt more classes. Thanks for your suggestion, we will further update a clean version.
For youtube-vos, the num_classes should be 40 (the background class is added by default in mmdetection 2.0).
https://github.com/Exely/masktrackrcnn-mmdet2.0/blob/83fcc2b0342f17e79d17d151ab8b1b11dc4359c8/configs/masktrackrcnn_ytvos/masktrackrcnn_r50_fpn_2x.py#L67
Why the num_classes is set to be
1
instead of41
or40
?