CODEWITHSUNDEEP
object-detectionintelyoloyolov5yolov4
M. Abdel-Ghani
M. Abdel-Ghani

Reputation: 205

Yolo version to run Oak D on CPU

In 2021 I was able to train a Yolov4-Tiny model to deploy on an Oak-D camera, and I was able to run the model on the camera on my PC using only CPU (I trained on Google Colab using GPU but for running on the camera CPU was enough). The FPS was very good for live detection.

Since then multiple Yolo versions were released and it seems that Yolov4-Tiny is not very much supported as before for running on the Oak-D. I am training another model now using Yolov5s, but running it on my CPU results in very low FPS (1-4 only) and this is causing alot of problems for the detection. I tried searching for the best Yolo version lighweight enough for running locally on CPU on the Oak-D but I cannot find anything. Has anyone been through this before? Which Yolo version would provide a similar performance to Yolov4-Tiny on CPU only?

Upvotes: 0

Views: 353

Answers (1)

M. Abdel-Ghani
M. Abdel-Ghani

Reputation: 205

Okay, coming in to answer this question after completing the project. It seems that running Yolov5 without optimisation requires GPU. So, in order to use CPU only, the model itself must be optimised further to run on CPU. Given that I was on a tight deadline, I did not have time to experiment with optimisation techniques, so instead I retrained my YOLOv5s model in Roboflow, and then loaded my model onto the OAK-D camera directly from the website. The model loaded this way was optimised for running on CPU only and I was able to achieve a much higher FPS than before. (Will definitely update this answer if I am able to use a manual way of optimising the model in the future)

Upvotes: 0

Related Questions