autonomousvision/unimatch

The time required for stereo matching

ynl-kusama opened this issue · 4 comments

ynl-kusama commented

Hello!
I am currently trying to integrate ROS system and unimatch, but the time required for stereo matching is taking more than 3 seconds and does not run as smoothly as the demo video.
My input image is 1920x1080 in size and I tried reducing the resolution, but it takes about the same amount of time.
Is there any point to increase the frame rate?

haofeixu commented

Hi, you can try to use smaller inference size for inference: https://github.com/autonomousvision/unimatch/blob/master/scripts/gmstereo_demo.sh#L7

ynl-kusama commented

Thank you for your reply!
It is true that if I reduce the inference_size, it takes less time to stereo matching.
For reference, could you please tell us the inference_size setting in your demo video?

haofeixu commented

Hi, for our demo on huggingface, we use a maximum inference size of 640x960: https://huggingface.co/spaces/haofeixu/unimatch/blob/main/app.py#L46
For inference on the full-resolution Middlebury dataset, we resize the images to 1024x1536: https://github.com/autonomousvision/unimatch/blob/master/scripts/gmstereo_demo.sh#L7

ynl-kusama commented

Thank you for your kind attention!