Closed Tortoise17 closed 2 years ago
@SilvioGiancola I have to ask for TempPooling in config. Like in CALF is
# Values of the K parameters (in seconds) in the context-aware loss
K_V2 = torch.FloatTensor([[-100, -98, -20, -40, -96, -5, -8, -93, -99, -31, -75, -10, -97, -75, -20, -84, -18], [-50, -49, -10, -20, -48, -3, -4, -46, -50, -15, -37, -
5, -49, -38, -10, -42, -9], [50, 49, 60, 10, 48, 3, 4, 46, 50, 15, 37, 5, 49, 38, 10, 42, 9], [100, 98, 90, 20, 96, 5, 8, 93, 99, 31, 75, 10, 97, 75, 20, 84, 18]]).cuda()
Can I have values for Temporary_Aware_Pooling NetVLAD++ for classes.py in config folder? I guess here I need your help. which values to be replaced for this model. Your kind help is highly appreciated.
Hi @Tortoise17 , You actually don't need K_V2
for NetVLAD++. All you need to set is the size of the temporal window you want to pool from. That temporal value is generic for all the actions, my model/code cannot handle multiple values per class.
Now I don't know why the pre-trained model does not work as well as expected in your own video, and if CALF works better, maybe it means that CALF transfers better to your video, I can't tell.
@SilvioGiancola Thankyou so much. Can you tell me as a guess about the temporal window size ? to decrease or increase?
at the moment, it is 15
The sensitivity analysis on SoccerNet was provided in the paper (Table 2): https://openaccess.thecvf.com/content/CVPR2021W/CVSports/papers/Giancola_Temporally-Aware_Feature_Pooling_for_Action_Spotting_in_Soccer_Broadcasts_CVPRW_2021_paper.pdf
Thank you so much.
I have to ask one question. I used NetVlad++ model and TempPooling inference results on my own data video test is that accuracy is less than CALF model json. All the events are less than value 0,25 in visibility. Could you give any hint why it could be. I guess that I need correct K_V2 values