Retraining with image size of 500x500 about squeezedet HOT 4 CLOSED

@xjoshramos The H and W should be approximately 1/16 of the input image size. However, due to integer rounding caused by padding/striding/down-sampling strategy, the exact number could be a bit different. Given input image size of 500x500, I would suggest starting with 32x32 and go through a few iterations of trial-and-error.

B is just the number of anchors for each grid center, so it really depends on how many objects can be inside one image, the range of object size, etc.

from squeezedet.

Thank you!

from squeezedet.

My training data is not all the same size. When training, does the the resizing affect my bounding box ground truth? Is the ground truth change according to the resize or do I need to handle that on my end?

from squeezedet.

@xjoshramos If you look at this line: https://github.com/BichenWuUCB/squeezeDet/blob/master/src/dataset/imdb.py#L189 You will see that bounding boxes are also scaled accordingly together with the image.

from squeezedet.