search

Cysu / dgd_person_reid

Domain Guided Dropout for Person Re-identification
http://arxiv.org/abs/1604.07528
231 stars 94 forks source link

My Python implementation performs much worse than the paper #33

Closed GBJim closed 7 years ago

GBJim commented 7 years ago

Hi all

I wrote a Python implementation covering the feature extraction and on VIPeR dataset. (The CMC function is imported from the CMC.py of this repo)

Using the pretrained JSTL+DGD mode provided by @Cysu, I only got 19.33% of Top 1 accuracy. It's far from the 37.7% on the paper.

I guess there are bugs in my script, but I couldn't find it.

Cysu commented 7 years ago

I'm not sure, but one possible reason is that we first resize the image to 160x64, then get center-crop of 144x56, instead of directly resizing it to 144x56.

GBJim commented 7 years ago

Hi @Cysu I forgot to transform RGB input into BGR format. The Top 1 accuracy reached 37.3% after adding transformer.set_channel_swap('data', (2,1,0))

I will also try the central cropping strategy you suggest.
Thank you!

Cysu commented 7 years ago

@GBJim Good to hear that! I didn't notice this point either.