NMS type error in execution of cpp example code of object detection failed on rk3588 based sbc with Hailo8L

I am trying to run objection detection with yolo11s (downloaded from official hailo model-explorer website) using c++ code provided at Hailo-Application-Code-Examples repository. Exactly, I am trying to run this example. It is getting compiled correctly, but when I try to run the executable file, I receive error as:
terminate called after throwing an instance of ‘std::invalid_argument’
what(): Output tensor yolov11s/yolov8_nms_postprocess is not an NMS type
Aborted (core dumped)

The same errror occured with yolov8s also.

exact execution command and output of terminal is as:

can someone help me to fix this?

I have already changed the architecture from x86_64 to aarch64, the issue still persists.

Hey @aamanraj121 ,

Could you please share the HEF model’s parsing details? The issue appears to be related to NMS post-processing. There might be differences between YOLOv8 and YOLOv11 output formats that are causing this. To better diagnose the problem, I’d appreciate more information about the model’s structure and configuration.

I have downloaded the model yolov11s.hef from hailo official model explorer website.
I think the issue is with the v20 software update. I downgraded driver and hailort package to v19 and the same model worked fine.

also the same model file worked well with python package pyhailort v20. It just creates issue in c++ with hailort v20

Hello,

I think @aamanraj121 is correct by saying that the problem lies in version 4.20.0 of the software. I have exactly the same error on my test application. The error happens on the official models yolov6n.hef and yolov8n.hef

On version 4.19.0 everything works on amd64 and arm64.

As a side note, while I appreciate very much async_detection_inference example for CPP, I must say it is not of very good quality :frowning:
The example is using thousands of futures (depending on number of frames) that are filling up the memory, which causes out of memory on bigger videos.

Cheers,

Chuck