statho
commented
3 months ago Thank you for appreciating our work!
-
The demo with videos requires 12.7 GB of GPU memory (due to tracking with 4D Humans).
-
The demo with images requires 14.4 GB of GPU memory, and uses several models:
- ViTDet : 2.8 GB
- ViTPose : 2.6 GB
- HMR 2.0 : 3 GB
- ScoreHMR (including PARE backbone) : 0.4 GB
Furthermore, the input images have high resolution (6K x 4K), and contain a lot of people.
To reduce the GPU memory usage, I suggest to:
- store detections from ViTDet and ViTPose first, and then run HMR 2.0 and ScoreHMR, OR
- run ScoreHMR on top of PARE regression, OR
- try the existing code on lower resolution images
Dear Author,
First of all, thank you for your excellent work. I encountered an Out of Memory (OOM) error while doing inference using the provided demo script. So I'm curious how much VRAM is needed for inference. I would be grateful if you could provide information on this.