search

rohit901 / cooperative-foundational-models

Official code for our paper "Enhancing Novel Object Detection via Cooperative Foundational Models"
https://rohit901.github.io/coop-foundation-models/
MIT License
49 stars 4 forks source link

Add inference code for user-defined data #2

Open chau25102001 opened 7 months ago

chau25102001 commented 7 months ago

Thanks for your great work. Can you add a script for model inference on a single input image? Thanks in advance

rohit901 commented 7 months ago

Hello,

Thank you for your interest in our work. Sure, I will try to add the script for it ASAP, and will let you know. Please stay tuned, and you may also opt in to "Watch" this repository in the meanwhile.

Thanks

ralpyna commented 7 months ago

Interested in your work. Thank you very much for it. It is the same request with @chau25102001. I would like to compare with results of other model as qualitative. Can I know your plans for this?

rohit901 commented 7 months ago

Thank you for your interest.

Before I make that, I would like to make a note that in the "novel object detection" setting (i.e. Table 1 results), we use MaskRCNN weights trained on COCO images/labels (80 objects), and GDINO-SwinT (trained on O365,GoldG,Cap4M).

Further, during inference we evaluate for 1203 LVIS classes using class synonym information along with the class names.

Thus, hope it will be okay for you guys if I push a script that works on your custom image data, but still you would be detecting the 1203 LVIS class objects in your image, and using the above two existing model weights itself.

If you would like more customization, I hope you would be able to make the necessary/required changes in the code? @ralpyna @chau25102001

rohit901 commented 7 months ago

@ralpyna, @chau25102001 I have added the code for inference on custom images. Please follow the instructions mentioned in the README to test it on your images.

I hope this will help you, and you would be able to modify or make other changes depending on your use case.

Please let me know whether I can close this issue or not.