search

bytedance / fc-clip

[NeurIPS 2023] This repo contains the code for our paper Convolutions Die Hard: Open-Vocabulary Segmentation with Single Frozen Convolutional CLIP
Apache License 2.0
276 stars 27 forks source link

Unable to reproduce performance, PQ=0 #31

Open zhaoyangwei123 opened 3 months ago

zhaoyangwei123 commented 3 months ago
Hello author, I attempted to replicate the performance of fcclip-r50 on the Coco dataset using the checkpoint you provided, but did not achieve satisfactory results. The performance of panoramic segmentation and semantic segmentation on the Coco dataset was always 0, but the results of instance segmentation can be replicated. What is the reason for this? [06/03 13:07:59 d2.evaluation.panoptic_evaluation]: Panoptic Evaluation Results: PQ SQ RQ #categories
All 0.000 0.000 0.000 133
Things 0.000 0.000 0.000 80
Stuff 0.000 0.000 0.000 53

[06/03 13:08:30 d2.evaluation.testing]: copypaste: AP,AP50,AP75,APs,APm,APl [06/03 13:08:30 d2.evaluation.testing]: copypaste: 40.6584,63.1976,43.7951,21.7050,44.9967,60.3094

[06/03 13:08:30 d2.evaluation.testing]: copypaste: Task: sem_seg [06/03 13:08:30 d2.evaluation.testing]: copypaste: mIoU,fwIoU,mACC,pACC [06/03 13:08:30 d2.evaluation.testing]: copypaste: 0.4226,1.6011,1.2336,2.6937

At the same time, I found that when I run Python datasets/prepare_coco_semantic-annos-from_panoptic-annos.py to extract semantic annotations from panoptic annotations, I get panopticsemseg {train,val}2017 All of them are as follows, is this because there is an issue with the annotations?

000000001503