• Hardware Platform Jetson
• DeepStream Version 7.0
• Issue Type Question
I am using DeepStream with YoloV10. I integrated YoloV10 using this set of instructions that I found in another post.
(Thank you for this contribution!)
In short I’m getting duplicate detections from YoloV10.
I’m getting much faster and better inference results from PeopleNet. YoloV8 was also much faster than PeopleNet as well but it was getting way too many false positives for people detections. YoloV10 seems to be getting the perf and accuracy I need except for one issue. For many of the people in the video I get two detections with bounding boxes that are largely the same but maybe vary by a few pixels from each other. I am just learning about the dropping of NMS in YoloV10 and the IOU settings in the downstream trackers… so I’m asking for guidance in where I should investigate first. I’m reluctant to add my own de-duping at the end of the pipeline because I don’t think I can get it right of which are false duplicates coming from YoloV10 vs cases where two people have passed closely behind one another. The other reason I don’t think I should be doing this externally is because I’m using downstream trackers (DeepSORT, etc), and if I wait for the inference results to prune duplicates, the trackers are still getting the duplicates.
Anyone else have any experience with YoloV10 and DeepStream and seeing this issue.