Single class output issue with quantized YOLO models on QCS6490

[freedomlyle]

Following up on a previous inquiry regarding custom dataset quantization (which was resolved by modifying the source code), a new issue has emerged.

It appears that models trained and submitted to the Hub for quantization consistently yield only a single output result when running on the GenKit QCS6490. Regardless of the dataset used, the detection output is strictly limited to objects with label 0. This behavior persists even when testing with the official YOLOv8 pre-trained model using the standard quantization workflow and COCO dataset; only label 0 (Person) is detected.

The environment is set up using the GStreamer sample application via API mapping on the board, following this tutorial:
https://docs.qualcomm.com/doc/80-70018-50SC/topic/gst-ai-object-detection.html

The corresponding source code path is:
gst-plugins-qti-oss-imsdk.lnx.2.0.0.r2-rel\gst-sample-apps\gst-ai-object-detection

After verifying the label files and configuration, a discrepancy was noted in the "constants" configuration compared to the tutorial.

The tutorial documentation shows 3 parameters for offsets and scales:
"constants": "YOLOv8,q-offsets=<21.0, 0.0, 0.0>,q-scales=<3.0546178817749023, 0.003793874057009816, 1.0>;"

However, the current parameters in my setup only contain 2 values:
YOLOv8,q-offsets=<0.0, 0.0>,q-scales=<2.4900553226470947, 0.00390625>;

Could you please advise on the cause of this discrepancy and how to fix the single-output issue?

[kory]

Hi @freedomlyle , I'm sorry we missed your question.

The challenge is in your quantization parameters. When you re-quantize the model, the inputs and outputs will have different quantization scales and offsets. You can take a look at your compile job; when you click on the "target model", you'll go to a webpage like this:

(THIS IS AN EXAMPLE, not your actual quantization parameters. You must look at your own compile job)
https://workbench.aihub.qualcomm.com/models/<model_id>

The new scales and offsets will be provided in the model input / output details sections, like above. You will need to update the scale and offset constants according to those values..