search

cvlab-columbia / pix2gestalt

Code for the paper "pix2gestalt: Amodal Segmentation by Synthesizing Wholes" (CVPR 2024)
https://gestalt.cs.columbia.edu/
Other
119 stars 8 forks source link

Testing script for amodal segmentation on COCOA #6

Closed trqminh closed 1 month ago

trqminh commented 1 month ago

Thank you for the interesting work! I would like to ask if you have the script to generate the result in Table 1 in the paper for amodal segmentation results.

egeozguroglu commented 1 month ago

Hi, we based our Amodal COCO and Amodal BSDS evaluation for amodal segmentation (Table 1) on the Amodal COCO API source code, using the inference functionality provided here.

Specifically, we sample multiple predictions from the diffusion process. Since our method is probabilistic, it encodes the inherent ambiguity and under-constrained nature of amodal completion.

For each predicted whole (amodal completion), we retrieve the corresponding binary amodal masks, then compute the best of 3 with majority voting — as reported in our paper.

Does this make sense? Please feel free to email me at ege.ozguroglu@columbia.edu for any further details!

trqminh commented 1 month ago

Hi @egeozguroglu , thanks for your response. It makes sense to me, I know how the procedure looks like. My question is whether you have the script of this procedure, meaning that I can run it and get the number that you reported in the paper.