Model in the FO1_VS_SAM3_DEMO

[Hickey8]

Great work! The demo in https://huggingface.co/spaces/P3ngLiu/FO1_VS_SAM3_DEMO is impressive. I'm wondering whether this demo uses VLM-FO1+SAM3 or VLM-FO1+UPN? The concept detection ability seems to inherit from SAM3 from my perspective, but the performance in the demo is much better than the provided script inference_with_sam3.py.

[P3ngLiu]

Hi!
This demo uses VLM-FO1+UPN for comparison. You can check the implementation of this demo at here. https://huggingface.co/spaces/P3ngLiu/FO1_VS_SAM3_DEMO/blob/main/demo/gradio_demo2.py

[Hickey8]

Got it, thanks for your reply! From my experiments, under open-vocabulary detection, the accuracy of VLM-FO1+UPN still seems slightly inferior to VLM-FO1+SAM3. While VLM-FO1+SAM3 tends to produce empty outputs more frequently in certain scenarios, its overall accuracy is higher. Have you conducted any quantitative or qualitative comparisons between these two approaches, or do you have relevant experience? Which method would you recommend for open-vocabulary detection?