search

ZhaochongAn / COSeg

[CVPR 2024] This repo contains the code for our paper: Rethinking Few-shot 3D Point Cloud Semantic Segmentation
69 stars 8 forks source link

Fail to reappear the mIoU reported in the paper. #1

Closed Pengz98 closed 6 months ago

Pengz98 commented 6 months ago

Hi! Thank you for your excellent work in few shot segmentation. I have tried to reappear the final result following the guideline, but didn't achieve the mIoU value reported in the paper (48.1 for 1-way 1-shot in S3DIS). To be specific, the mIoU value reaches the maximum (45.38) in Epoch 13, and remains around 40 till the Epoch 100. I am not sure if there is any detail I didn't aware of. It will be helpful if you could offer the well-trained weight of the few-shot training phase.

ZhaochongAn commented 6 months ago

Hello! Thank you for the interests in our work. The training process does have some variance. And sometimes it could happen that the best-performing model may be obtained from the early checkpoints. I have released the model weights. Hope it can be helpful.