Jetson TX2 NX Jetpack 4.6 deepstream 6.0.1 CUDA 10.2 cudnn 8.2 TensorRT 8.0 Yolo V5 6.1
When I finished deploying YoloV5, it worked poorly, detecting a video that yielded very few targets（they are almost the same class） and all of them were wrong.The model in pytorch can work well.
The engine file is in FP32 model,nms-iou-threshold=0.01 and pre-cluster-threshold=0.01.
Hello @1105781123 Are you running opensource Yolov5 or TAO YoloV5? Is DeepStream involved in your test?
I am running opensource Yolov5, DeepStream is involved in your test(run deepstream app ).
Now I choose another model(more epochs in training).That model can work batter,most targets can be detected.But there is also something wrong:There is almost always a misidentified target in the upper left corner of the window. They are all gray labels that do not show confidence. (Other labels are red, orange, etc.)
When I set pre-cluster-threshold=0.95，this problem has been greatly alleviated, but it still occurs occasionally
Do you have test result without using TensorRT/DeepStream? need to narrow down if it is related with precision of the model or not.
I tested the model in pytorch , it can detect the targets.Detecting a viedo (1000 - 1500 targets), about 3 of the targets were mistakenly identified.
This looks like a TensorRT related issue. We will move this post to the TensorRT forum.