diffrent result on HF and my PC

SHI-Labs / OneFormer

OneFormer: One Transformer to Rule Universal Image Segmentation, arxiv 2022 / CVPR 2023

https://praeclarumjj3.github.io/oneformer

MIT License

1.39k stars 129 forks source link

diffrent result on HF and my PC #114

Open xuyuxiu83 opened 2 months ago

xuyuxiu83 commented 2 months ago

Thanks for ur excellent work! I tried the same image and the same backbone on Hugging Face and on my computer separately, but I got different results. It's very clear that the results on my computer cannot recognize water. 150trail-3_02001_semantic_inference water_seg

xuyuxiu83 commented 2 months ago

How should I do to improve the performance? The above results are from using DINAT as the backbone, trained on the COCO dataset, for making predictions.

vietpho commented 2 months ago

+1 I'm experiencing the same issue. The Hugging Face demo performs significantly better. I compared it locally with the same model and the same images without modifying anything, yet the performance discrepancy remains.

praeclarumjj3 commented 2 months ago

please check the comment in #115