Reduced model for realtime applications?

apple / ml-depth-pro

Depth Pro: Sharp Monocular Metric Depth in Less Than a Second.

Other

3.29k stars 205 forks source link

Reduced model for realtime applications? #24

Open Mh-Magdy opened 1 week ago

Mh-Magdy commented 1 week ago

I ran the model on a high-quality image and it took about 8 Seconds to infer the depth. I'm wondering if there is a smaller model suitable for real-time application or any other suitable edits for this use case? Device: NVIDIA RTX 3060

YCHAOSK commented 1 week ago

same issue

harism commented 3 days ago

I've reached 1-2 seconds with these changes running Depth Pro 1024x1024 float16 on MacBook Pro M2 Neural Engine; https://github.com/apple/ml-depth-pro/pull/45 - I believe you can do something very similar with normal execution also and reduce the model inference size.