search

chenhaomingbob / CSC

[CVPR 2024] This is official implementation of our CVPR 2024 paper "Building a Strong Pre-Training Baseline for Universal 3D Large-Scale Perception" https://arxiv.org/abs/2405.07201
Other
11 stars 0 forks source link

How to fix the superpixel number from the SAM segmentation results #1

Open cppcute-pm opened 2 months ago

cppcute-pm commented 2 months ago

Hello, thanks for the amazing work you've done! I've read the released code and I noticed that there are no superpixel number information in the proprocess code for SAM related to semantic segmentation. But the code of lightning trainer obviously need the superpixel_size to make the sparse matrix. My current thought is that you use the segmentor in skimage to generate the superpixel and use the SAM to generate the semantic masks. Is that right or there are some other details?

chenhaomingbob commented 2 months ago

Thanks for your attention. We directly use DINOv2 to produce superpixels, and do not employ slic or SAM.

cppcute-pm commented 2 months ago

Thanks for your reply! I'm interested in how you get the superpixels_size fixed when using the DINOv2. I guess there must exists important processing steps to achieve it and I'm very curious about the details. Could you provide the code related to it or describe the process pipeline in detail? My email address is 2233589@tongji.edu.cn.