Problem with the superblob config file for yolov6r3 conversion

AyushmanChoudhuri

Hello!

I have been converting custom yolov6r3 (nano) models to superblobs using the model converter api using the code snippet blow:

converted_model = convert.RVC2(api_key=LUXONIS_HUB_API_KEY,

path="best_ckpt.pt",

name="YOLOv6",

description_short="superblob",

yolo_version="yolov6r3",

yolo_input_shape="512 512",

superblob=True,

yolo_class_names=labels,

tasks=["OBJECT_DETECTION"],

license_type="MIT",

is_public=False

)

However the config file inside the NNarchive zip file shows that:

The model is yolov6r2
The confidence threshold is 0.5 ( I would like this to be 0.05)
The reverse-channels is set to false.

Is the config file actually used for the superblob inference. Because I do see output confidences of 0.1, even though the config file mentions the threshold to be 0.5.

Is there a way to update the config file parameters, incase it is used for the inference.

KlemenSkrlj

Hi @AyushmanChoudhuri ,
Let me give a bit more context to the values you see in the NNArchive config file:

The model is yolov6r2

This is expected because YOLOV6R2 and R3 have the same type of postprocessing so we just reused the earlier version in the postprocessing code. Meaning that although the model is R3 its postprocessing follows the R2 path and the outputs are still correct.

The confidence threshold is 0.5 ( I would like this to be 0.05)

The 0.5 is the default confidence we set when converting YOLO models. You can change the confidence at runtime though through the API. If you are using depthai.node.DetectionNetwork you can check out the usage here (detection.setConfidenceThreshold(0.05)) or if you are using the depthai_nodes.node.ParsingNeuralNetwork check out the docs here.
In either case you should not be able to see detections with 0.1 confidence with 0.5 confidence threshold is set. So if you have a MRE of this please do share so we can investigate it further.

The reverse-channels is set to false

Here maybe more important flag is the dai_type. In your case this is set to BGR888p which means that the model is exported so that it expects BGR [0,255] planer (CHW) inputs. But the pytorch YOLO model was trained on RGB images. So what reverse_channels=False in this context means is that this color channel reverse from BGR (DepthAI default) to RGB (what model was trained on) is already built into the exported model and you don't have to take care of it manually. Essentially when working with NNArchive DepthAI is able to read all the preprocessing information from it automatically and configure the inputs correctly so that you don't have to think about this.

Hope this helps,
Klemen