Obtain final action scores of two or four-stream training

[anhvu2501]

Thank you for your wonderful work! I would like to ask if you can let me know how you obtained the final action scores (final top-1 accuracy) of the two or four-stream network. As I have read in the MS-AAGCN paper, the author says that the softmax scores of the four streams are fused using weighted summation. Which weights I should assign for each stream to receive an approximate scores as your official implementation?

[wujiangru]

Hello, have you found a way to fuse the joints and bones?

[anhvu2501]

Hi, I tried by following the author's approach that is the softmax scores of the multi-stream are fused using weighted summation. Then, I just assigned weights for joint and bone modality, based on the benchmark we are evaluating (For instance, if we evaluate on XSub benchmark, I assigned weight for bone modality higher than for joint, vice versa) Do you think whether this approach is correct or not?

[wujiangru]

Thank you for your reply, I think this is fine. I am currently allocating on a 1:1 ratio. I will use your suggestion to make adjustments.

[kennymckormick]

It's fine I think. You just need to use the weighted summation of scores for different modalities (joint, bone, etc.) as the fused prediction.

[YYDSbiubiu]

Hello, I would like to ask you a few questions. Can you tell me how to get the two-stream accuracy? Could you please share the code with me? :)

[masOskar]

please share the code