Crazy yolov4-tiny detections from depthai-python examples

elitearmedforce

I was trying out the demo code of spatial detection with yolov4-tiny in depthai-python examples for OAK-D (spatial_tiny_yolo.py). Most of the time detections detect more than a single car although OAK-D camera is pointed at a single car image, same goes for people too. I am not sure why it is the case. Anyway, to solve this issue, please help.
Screenshot attached:

erik

Hello elitearmedforce ,
I believe setting appropriate IOU threshold fixes this issue. Maybe try setting a larger value (closer to 1.0):

yoloDet = pipeline.create(dai.node.YoloDetectionNetwork)
yoloDet.setIouThreshold(0.5) # Try setting higher value

Thanks, Erik

elitearmedforce

erik Thank you for the reply. Interesting that I didn't get any email notifications.
I tried messing with that value, and it got worst(screenshot). Then I tried messing with yoloDet.setConfidenceThreshold(0.8) [changed 0.5→0.8] and that worked out for me. Thanks again this worked for me(setConfidenceThreshold). From my understanding IOU, higher of it should give better results, whereas in this code the lesser it is the better (correct me if i am wrong)

with yoloDet.setIouThreshold(0.8) & yoloDet.setConfidenceThreshold(0.5)

with yoloDet.setIouThreshold(0.1) & yoloDet.setConfidenceThreshold(0.8)

Code Snippet: Github - spatial_tiny_yolo.py OR Pastebin

erik

Hi elitearmedforce , My bad, I thought the higher IoU value would provide more filtering, but it has the opposite effect. Great that you figured it out and thanks for the info!