search

MCG-NJU / RTD-Action

[ICCV 2021] Relaxed Transformer Decoders for Direct Action Proposal Generation
Apache License 2.0
86 stars 8 forks source link

One question about the the boundary attentive module #9

Closed 3huo closed 2 years ago

3huo commented 3 years ago

Hi~ Does the input feature and start/end scores is pretrained in advance?

3huo commented 3 years ago

And I also concern that if my start and end scores are predicted very correctly, then the center feature after multiplication could go to zero. can it still very effective for proposal generation?

JackyTown commented 3 years ago

No, input features and start/end scores are directly trained on target datasets.

Actually, start and end scores cannot be predicted very correctly. (If boundary probability of center features are zero, TAPG will be totally solved.) For more details, you can refer to Appendices E.

3huo commented 3 years ago

Oh~, Thank you for your replying! so... the input features training process and the encode/decode process are trained separately?

sparkstj commented 2 years ago

Yes