How was your models trained?

[joeyhng]

Does your released model trained using the temporal annotations, or just the video level labels? Thanks.

[yjxiong]

The challenge models are trained with activity instances clipped according to the temporal annotations.

If you are looking for an approach using only video level label, please check out our CVPR17 paper UntrimmedNets for Weakly Supervised Action Recognition and Detection.

[joeyhng]

Thanks.