For the target detection task, in the pytorch (yolov8) model, I added two additional branches based on the original obj and cls branches, one with output channel 1 and the other with output channel 51. The order of vsstream output from the original yolov8 hef is (80, 80, 64), (80, 80, 20), (40, 40, 64), (40, 40, 20), (20,20, 64), (20, 20, 20). After adding two branches, the output index changes.
Hey @jiqimaohw
Lets start by doing the following :
Update Model Outputs:
- Ensure new branches are integrated into the PyTorch YOLOv8 model and marked as outputs.
Inspect HEF:
- Use the
parse-hef
tool to check the output vstream order.
Update Application Code:
- Modify inference code to match the new output indices using HailoRT