search

nasir6 / zero_shot_detection

63 stars 18 forks source link

How to get the Qualitative Results. #13

Open Programmergg opened 3 years ago

Programmergg commented 3 years ago

Hi, thank you for providing the excellent repository! After reproducing your results, I want to get the detection boxes for zero-shot detection or generalized zero-shot detection on the images. That is how to get your Qualitative Results. Look forward to your reply.

Dai1029 commented 2 years ago

Maybe a little sudden. I'm having some trouble reproducing (only achieving map=8%). I want to ask you if there are any areas or parameters you need to pay attention to. I hope to get your reply.

Programmergg commented 2 years ago

Maybe a little sudden. I'm having some trouble reproducing (only achieving map=8%). I want to ask you if there are any areas or parameters you need to pay attention to. I hope to get your reply.

Hi, I would like to get the information about your conda environment. From my experience, the ZSD code based on mmdet1.x is only able to get better results with older conda environments, e.g. , torch1.1+cuda9-cuda10. I have tried migrating this code to mmdet2.x or deploying mmdet1.x on newer hardwares(e.g. V100 or A100) and unfortunately they both give poor results, similar to yours. Of course, it is possible that my parameters need to be adjusted, but I didn't look into it further. At the time of my original reproduction, I was using an environment of torch 1.1.0 + cuda 10.2.0 and two RTX 2080Ti or RTX TITAN graphics cards. I would recommend you to run it with RTX2080Ti + 440.44 graphics card driver, it gets more stable and easy to configure results.

Dai1029 commented 2 years ago

I moved the environment to mmdetection2.x, which is probably why I couldn't reproduce it.Thank you very much for your help.