Spaces:

---

sergiopaniego

/

vlm_object_understanding

like 89

Running on Zero

Hey @sergiopaniego 👋

We just released a new version of Supervision (v0.26.0) with even more support for Vision Language Models (VLMs)! 🧠🖼️
🔗 Release: https://github.com/roboflow/supervision/releases/tag/0.26.0

This PR builds on that by adding VLM-related improvements, especially since the original code is based on Qwen2.5-VL and Moondream2 models. I’ve also used the new from_vlm feature to simplify bounding boxes and annotations to making the code more cleaner and intuitive. ✨
And lastly I removed all of the unused imports to keep code and space cleaner.

I hope you like it.

Thanks a lot!!!

Just one concern. In a previous PR, we moved the @GPU to improve the accuracy when considering inference time. I can see that in this PR, that idea is reverted.
Is this something that you've benchmarked and works better? :)

I was working on locally and I removed space and GPU tag to test it. I accidently put the wrong place. But after I check the PR you linked me I put it back properly for both inference funcs

Thanks!
In case you're interested, HF has a PRO mode (https://huggingface.co/pro) so you can directly code remotely with Dev Mode (https://x.com/SergioPaniego/status/1937197940621115657)