YOLOv6 .pt to .blob Conversion Not Working

jakaskerl I used the Roboflow YOLOv6 training notebook found in this article. It seems to be a bit different than the one you provided, I'll give the one you provided a go in the meantime. Thanks!

jakaskerl Nevermind, these seem to be the same. The only difference in the one that you linked is that it uses older pip libraries and that seems to make it unusable as the libraries don't like each other. Even after fixing the libraries, it seems like the older YOLOv6 doesn't work seamlessly.

TheOracle Thanks for your suggestion, but unfortunately that's the website I'm already trying. I think you're right though, the AxiosError seems to be something wrong with the tool :/ if anyone knows an alternative way of converting my .pt to a group of .bin, .json, .xml, .blob, and .onnx files, that would be much appreciated!

JanCuhel sounds good. Thanks.

JanCuhel Hey Jan, sorry to bother you, but my project is on a tight schedule, any updates?

JanCuhel Thanks a lot Jan, looking forward to it

JanCuhel Hi Jan, thanks a lot for your help. However, here is my script:

from depthai_sdk import OakCamera, ArgsParser
import argparse

latest_z_value = None  # Store the most recent Z distance

def process_detections(packet):
    global latest_z_value

    if packet is None or not hasattr(packet, 'detections'):
        print("No detections found.")
        return  

    for det in packet.detections:
        if hasattr(det, 'img_detection') and hasattr(det.img_detection, 'spatialCoordinates'):
            latest_z_value = det.img_detection.spatialCoordinates.z / 1000
            print(f"Drone detected {latest_z_value:.2f}m away.")
        else:
            print("No spatial data available.")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("-conf", "--config", help="Trained YOLO json config path", default='models/v1/16shaves/droneDetection_v1.json', type=str)
    args = ArgsParser.parseArgs(parser)

    with OakCamera(args=args) as oak:
        color = oak.create_camera('color')
        nn = oak.create_nn(args['config'], color, nn_type='yolo', spatial=True)
        nn.config_nn(resize_mode='stretch')
        
        oak.callback(nn, process_detections)
        oak.visualize(nn, fps=True)
        oak.start(blocking=True)

if __name__ == "__main__":
    main()

As you can see, I am taking in the .json file and not the .blob. This is because the tools.luxonis.com tool would output a .bin, .json, .xml, .blob, and a .onnx file. Is there any chance that you can help with generating those.

Thanks!

JanCuhel Hi Jan, super helpful.

I downloaded the archive files and ran it on my script and noticed that the 640352 and 640384 are not the best at drone detection at a distance, but they run at 26-30fps. However, 1280_704 detects object at a distance, which is why I trained it on a larger resolution in the first place, but the fps is 4-6 :/

Do you know why there is a significant performance reduction? Is this related to the # of shaves?