Question about the embeddings
GriesserP opened this issue · comments
GriesserP commented
Hello,
Thank you very much for your contributionS and the realease of the embeddings.
I have some questions about them.
- Are the embeddings of size 8576 a concatenation of tpn_50, gta,vtn,ircsn,i3d-slow in this order ? So if I retrieve the first 2048 dimension ,do I end up with the tpn_50 features ?
- If yes to the previous question, we don't have access to the tpn_101 featuers right ?
- I am in fact trying to reproduce your idea of feature combination for an other dataset of football videos and I think that the process isn't trivial at all since it requires to understand and adapt the code of 5 other papers (one for each tpn,gta,vta,slowfast and ircsn). So could you provide the code to reproduce these embeddings ?
Cordially, Paul
kl2005ad commented
The order is tpn_101(2048d), gta(2048d), vtn(384d), ircsn(2048d), i3d-slow(2048d).
We cannot release the code for now. Please refer to each paper's released code or third party's implementations.
GriesserP commented
Ok I understand, thank you for the help!