chuanqi305
256x256 input size and question about pre-processing of images
Hi all!
When I am training mobilenet-ssd I got two problems.
Question1: I change the input size from 300x300 to 256x256. So the size of six mbox change the size from 19x19, 10x10, 5x5, 3x3, 2x2, 1x1 to 16x16, 8x8, 4x4, 2x2, 2x2, 1x1. I am not quite sure if it is reasonable because the loss keeps around 4 to 6 when the iteration goes to 20k. So I wonder could anyone give me some advic on it please? Thank you!
Question2: The input JPEG images are in different input size, when I convert them to lmdb dataset I used resize. Does it matter to xml file of the coordinate for the object? And if I use crop to pre-process the image when training, maybe some objects are not in the cropped image but the coordinate of this object is still kept in xml files, does it matter?
Thank you all!
