search

MKFMIKU / Instance-Shadow-Diffusion

Other
31 stars 2 forks source link

Inconsistency with the latent feature map when compared with the results shown in the paper #7

Open atsatvik opened 3 weeks ago

atsatvik commented 3 weeks ago

These are the results I got after I ran the test script with the pre-trained weights, I'm confused because they aren't anything like what's shown in the paper. I'm probably missing something. Any help on how I can get the correct latent feature map is highly appreciated. I'm trying to adopt some of the techniques you used in my own research and I wanted to experiment with the latent feature maps. Please let me know I can get it. Thanks!

exp_000000

This is what the latent feature map should look like, taken from the original paper. image

MKFMIKU commented 3 weeks ago

Hi @atsatvik , could you give more details about how do you get the latent feature map?

atsatvik commented 3 weeks ago

Hi @MKFMIKU, thanks for the reply.

I'm just running the shadow_aistd_test.py script on my single GPU machine using python shadow_aistd_test.py with the accelerator.load_state set to the path of the pretrained weights that you provided. The final results are saved in the experiments folder.

Edit: whats confusing is that even with this latent feature map, its doing shadow removal properly (for this case atleast)

atsatvik commented 3 weeks ago

Here are a few more: exp_000004 exp_000005 exp_000006