Questions about DSSD

[youngwanLEE]

Hi, Wei.

Recently, I have re-implemented DSSD based on your SSD code.

I just followed your arxiv paper.

DSSD used 321 resized input and was down-sampled to 161->80->40(a)->20(b)->10(c)->5(d)->3(e)->2(f).

So, when I tried to deconvolve the feature map(f) with kernel_size=4, stride=2, and pad=1 and eltwise sum with (e), shape miss match error occurred. Because the size of deconvolved feature map(f) is 4.

My question is how to solve this problem and how to configure deconvolution layer.

Thanks

[weiliu89]

Check BVLC#2213

[roipony]

@youngwanLEE
Can you share the DSSD code/configuration?

Thanks

[youngwanLEE]

@roipony I couldn't finish this re-implementation, yet. When I finish this project, I'll share it