search

happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
419 stars 77 forks source link

Training without gt_segments, gt_labels. #78

Closed grayfactory closed 1 year ago

grayfactory commented 1 year ago

Hi, thanks for great work! I have question about to training model on dataset which without segment and label informations.

I expected that if the model could be trained by the THUMOS14 dataset, it would be equally possible to train with without annotations. but I found in meta_archs.py

 assert video_list[0]['segments'] is not None, "GT action labels does not exist"
 assert video_list[0]['labels'] is not None, "GT action labels does not exist"

Is it possible to train with data without label and segment information? Looking forward to your reply. Thank you.

happyharrycn commented 1 year ago

This is not possible. Our model assumes a fully supervised setting, where action annotations (labels and onsets/offsets) are provided during training. Indeed, our work and prior works considers a subset of THUMOS14 dataset with temporal annotations. More details regarding the datasets and training schemes can be found in the paper.

grayfactory commented 1 year ago

Thank you for your quick reply. I get the point!