Closed think-high closed 5 years ago
Okay, It throws an error if I give a different crop size, which is what I expected.
1). So, if I need to let it take the entire width and height then I am guessing that I need to change the scale_h and scale_w to 112, right? Assuming that it resizes the image to scale_h x scale_w and then crops the crop-size from it.. 2). And if I don't want to scale it down but rather change the crop size to something bigger (Eg: X), then I need to train the entire model from scratch on Sports1m or Kinetics and set the crop-size to X there, right? 3). Also, if you can put some light on why you used 112 crop size and scale_h x scale_w as 128x171, then that would be great. Was it to reduce the computation time or you also found it giving better results?
Sorry for this bigger question asking a lot of details at once.
TIA.
Currently it works with crop_size 112x112 and length clip is a multiplier of 8. If you like to do something like FCN, you need to modify the final global pooling kernel size.
Hi,
I am using the pre-trained models provided in this repo, to extract features. I have a couple of questions regarding the crop-size argument:
1). If I let the crop-size be the default (i.e 112) then it crops the video to that size instead of resizing, right? 2). What If I don't want to crop the video because my video's dimension is way larger. Then what happens if I pass a different value to the crop-size during feature extraction?
TIA, Rahul