search

AndreyGermanov / yolov8_inference_video_javascript

Detect objects in images in a web browser using the YOLOv8 neural network
GNU General Public License v3.0
26 stars 11 forks source link

Inference FPS on video? #1

Closed sanjeevnara7 closed 1 year ago

sanjeevnara7 commented 1 year ago

Great work on this repo! I would like to know what FPS you are able to achieve during inference on the video?

My experiments with ORT web only give me 500ms/per image with the YOLOv8 which is really low (like 2fps?). I am interested in achieving 20+ fps on an input mp4 video, do you have any ideas?

Best regards

AndreyGermanov commented 1 year ago

I have similar results. Sometimes 400ms per image, sometimes 934ms. It depends on CPU power. Not sure that without GPU is possible to achieve much better results than you achieved.

Currently, I am discovering DeepSparse technology, which is aimed to run models on CPU with GPU speed.(https://docs.neuralmagic.com/).

Will share results in case of success.

sanjeevnara7 commented 1 year ago

Hi @AndreyGermanov thanks for your response. Using TensorflowJS with YOLOv8, I'm actually able to get 17fps with WebGL, while ORT web still only gives me 2fps (I also checked if WebGL is being used in the backend). Not sure if this is an issue with the library, since I'm running on an m1 mac.

Deepsparse gives me 15ms inference per image on an 8-core CPU which is very promising!

Best regards