How was Figure 4(a) produced?

[XuM007]

Thanks for your wonderful work. I am curious about how fig4a in the article was drawn.

# rgb_cams_ = (rgb_cams_ - self.mean.to(device)) / self.std.to(device) # B*S,3,H,W
# feat_cams_ = self.encoder(rgb_cams_) # B*S,latent_dim,H/8,W/8
feat_cams_ = rgb_cams_

I simply set feat_cams_ to rgb_cams_ in the mvdet code, and I can get a reasonable single-view project visualization result, but I can't get a reasonable multi-view fusion result. (Tried on the multiviewx dataset)

pro_img_all = feat_mems[0].permute(0, 2, 3, 1).cpu().numpy() # 6,3,160,250 -> 6,160,250,3
pro_img_all = np.mean(pro_img_all, axis=0) # 160,250,3

Can you provide ideas for making this figure? I think this is a good way to verify whether the mapping process is correct.

[2059189826]

@XuM007 Hello, I'm also very interested in creating this diagram. Do you know how to draw it? I look forward to your reply, thank you.