Open msarmiento3 opened 3 years ago
Hi @msarmiento3 , this is just for preprocessing the data, to have faster decoding and subsequent smaller-distance bilinear up/downsampling. Imagine if training videos would be of a very large resolution, the scale-jitter would cost a significant amount of time. The model-specific resolution is not related to this.
The input size of the frames for the X3D_L configuration is a crop of 312x312 from a frame with short size between 356 and 446. However in the repo you mention that all the kinetics videos are resized to have a short size of 256. Are this resized videos used for this configuration? or do you use videos resized to have short size of 356? I guess that there could be a performance drop when training with the smaller videos. Thank you!