YOLOv5 & YOLOv8 support for the OpenVINO Detector #5523
Conversation
✅ Deploy Preview for frigate-docs canceled.
|
|
Thanks y'all. I am considering adding some Google Collab notebook links in the docs for people to be able to retrieve and supply their own YOLO models to Frigate. I suspect there will be lots of questions about that, so a ready-made notebook would alleviate lots of those questions. Any thoughts? |
Could probably be written as a guide. I'd do that as a separate PR though |
| (object_detected[1] - (object_detected[3] / 2)) | ||
| / self.h, # y_min | ||
| (object_detected[0] - (object_detected[2] / 2)) | ||
| / self.w, # x_min | ||
| (object_detected[1] + (object_detected[3] / 2)) | ||
| / self.h, # y_max | ||
| (object_detected[0] + (object_detected[2] / 2)) | ||
| / self.w, # x_max |
There was a problem hiding this comment.
This portion seems to be common with all three yolo models. Could refactor to a utility function.
There was a problem hiding this comment.
Take a look at the latest commit.
When rethinking and going over the existing code, I found that the SSD post-processing might have a bug, where if a detection with less than 0.1 score is encountered before all 20 spots in the detections array is filled up, the for loop will exit prematurely without going through the rest of the inference results, even though there may still be other inference results with confidence scores higher than 0.1. FYI.
|
@NickM-27 the checks are stuck for some reason. Is there a way to kick them off again? |
I have added OpenVINO support for both YOLOv5 and YOLOv8 models. This diff adds on top of the previous work where YOLOX support was added. These are state of the art YOLO models, and I have observed incredible performance with them. Looking forward to seeing others try them out.