SilvioGiancola / SoccerNet-code

SoccerNet: A Scalable Dataset for Action Spotting in Soccer Videos
107 stars 25 forks source link

How can I get labels_Delta_.json? #17

Closed Dondkeman closed 4 years ago

Dondkeman commented 4 years ago

First, thank you for providing useful dataset and code.

I have tried to get more accurately chunk classification. and now, I try to evaluate new trained model for spotting task. but, I couldn't correctly running evaluation code.

Generation of predictions00(Argmax/NMS/Center_00.json was success. but, I can't running 'get_detection_performance_spotting', because labelsDelta.json files.

I read this comment https://github.com/SilvioGiancola/SoccerNet-code/issues/6#issuecomment-613969675 from closed issues page.

How can I get(or generate?) labelsDelta.json ??

Thank you.

SilvioGiancola commented 4 years ago

You can copy the structure from your prediction in the json format, and create the labelsDelta.json from Labels.json for each delta. Basically, Labels.json is for spotting (single timestamp) while labelsDelta.json is for activities (start/end timestamp). Then I used the code from ActivityNet to estimate the mAP with tIoU>0 for each deltas.

Alternatively, look at this repo: https://github.com/cioppaanthony/context-aware-loss there is a function for the spotting evaluation (Average-mAP) that avoid this transformation from spotting to activities.

Dondkeman commented 4 years ago

That was very helpful, thank you.