Explanation of this implementation

[Guptajakala]

Hi, thanks for the great repo! I'm interested in finetuning the model for my research and I'm trying to understand your implementation here:

r2d2/tools/dataloader.py

Line 119 in 740f6fb

scale = np.sqrt(np.clip(np.abs(dx[1]*dy[0] - dx[0]*dy[1]), 1e-16, 1e16))

1. what is this line doing?

2. From the comments above, it says "applying a median filter" but I didn't see median filter afterwards. Is this comment deprecated?

3. r2d2/tools/dataloader.py

   Line 145 in 740f6fb

   for _ in range(50*self.n_samples):

   Seems here you are doing some window selection based on best optical flow quality. Is this part necessary? Does it make big difference compared with random sampling a valid window inside the image region (without score comparison)?