Decoding tubes about moc-detector HOT 1 CLOSED

Oh, in my opinion the center frame doesn't need movement estimation. But we need to recover the numerical difference from discretization. (in ~/src/datasets/sample/sampler.py Line#L117)

So the movement contains a local offset for each frame, and for key frame (i=K//2), movement only represents the local offset.

On the other hand, wh prediction should be gathered from current frame's action center. And the range of numerical difference is (0, 1). So indexes of tube centers at key frame (K//2) are not modified by "mov" data. (Line 58-59)

In short, xs_all and ys_all in lines 58-59 are used to gather wh prediction, but codes in lines 80-83 are proposed to count predict boxes.

from moc-detector.