Ssd.pytorch: How to derive the 'steps' in config.py?
The 'steps' in config.py is [8, 16, 32, 64, 100, 300]. I am just wondering how to derive these numbers? I have read the papers which says 'f_k is the size of kth square feature maps', though I cannot relate it with the numbers you got. Thanks.
yxchng
All 4 comments
I think they are the total stride that branches out for each classifier, say the first 8 corresponds to 300 / 8 = 38 (x 38) feature map. Therefore, my doubt is that should the last two in the list be, 128 and 256?
hli2020
on 12 Sep 2017
Got it. Since the final feature map size is 3 x 3 and 1 x 1, the input is 300 x 300; therefore the stride should be 100 (approximate value to replace 128), 300 (approximate value to replace 256). Inspired by @chizhizhen
hli2020
on 12 Sep 2017
In fact, I wonder why not just use the 'feature_maps' : [38, 19, 10, 5, 3, 1] defined in config.py?
dingjiansw101
on 6 Feb 2018
The step is receptive field?
nationalflag
on 9 Oct 2018
Related issues
chienyiwang
路
3Comments
Elyorcv
路
4Comments
sxyxf66
路
4Comments
kouyichi
路
5Comments
DHZS
路
5Comments
Most helpful comment
I think they are the total stride that branches out for each classifier, say the first 8 corresponds to 300 / 8 = 38 (x 38) feature map. Therefore, my doubt is that should the last two in the list be, 128 and 256?