L

li_you_chen

@li_you_chen this isn't reproducible, as I do not have your model. Please upload whole zip to drive and share the link.

@li_you_chen please provide minimal repro example that's formatted correctly (use ```)

Hi @li_you_chen ,
Yep, that's possible. So you can have pipeline like this:

@li_you_chen this is because you're doing inference on 320x320 frames (1:1 aspect ratio) and drawing bounding boxes on full 640x400 frame (not 1:1 aspect ratio). Perhaps crop 640x400 into 400x400 (so 120 pixels on left/right), or just visualise boxes on imagemanip output frame (320x320).

@li_you_chen ,

objectTracker.passthroughTrackerFrame.link(trackerOut.input)
objectTracker.out.link(trackerOut.input)

So you are sending 2 different messages to the same queue - likely this is the problem.

Hi @li_you_chen
If you get no tracklets, the problem is likely the detection part. Please create another XLINKOUT node and pass detectionNetwork.out to it, and print it on the host. This will tell you if detections are made. Then we can debug further.

Thanks,
Jaka

Hi @li_you_chen
We have some here but we ditched MBNet a while ago and mostly use YOLO now. If you wish you can train YOLO as well.

Thanks,
Jaka

Hi @li_you_chen
Yes, of course you can. The model may be different but the on-device decoding is done in a way that ensures the output structure is the same, so they can both be linked to an object tracker.

If you train your own mobilenet model, the example you sent should be working without any changes (well, except for the model path).

Thanks,
Jaka

Hi @li_you_chen

li_you_chen full of wrong boundingbox , even I raise threshold

This suggest a problem with the model. Did you test it off depthai?

Thanks,
Jaka

Hi @li_you_chen
This is the exact same script as tiny_yolo in the examples. I only changed the model to catdog and the label names respectively. Can you recheck please?

Thanks
Jaka

Hi @li_you_chen
No, it's yolo. I'm just saying you can not use a yolo blob inside MobileNetDetectionNetwork.

#!/usr/bin/env python3

"""
The code is the same as for Tiny Yolo V3 and V4, the only difference is the blob file
- Tiny YOLOv3: https://github.com/david8862/keras-YOLOv3-model-set
- Tiny YOLOv4: https://github.com/TNTWEN/OpenVINO-YOLOV4
"""

from pathlib import Path
import sys
import cv2
import depthai as dai
import numpy as np
import time

# tiny yolo v4 label texts
labelMap = [
    "cat", "dog"
]

syncNN = True

# Create pipeline
pipeline = dai.Pipeline()

# Define sources and outputs
camRgb = pipeline.create(dai.node.ColorCamera)
detectionNetwork = pipeline.create(dai.node.YoloDetectionNetwork)
xoutRgb = pipeline.create(dai.node.XLinkOut)
nnOut = pipeline.create(dai.node.XLinkOut)

xoutRgb.setStreamName("rgb")
nnOut.setStreamName("nn")

# Properties
camRgb.setPreviewSize(320, 320)
camRgb.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
camRgb.setInterleaved(False)
camRgb.setColorOrder(dai.ColorCameraProperties.ColorOrder.RGB)
camRgb.setFps(40)

# Network specific settings
detectionNetwork.setConfidenceThreshold(0.5)
detectionNetwork.setNumClasses(2)
detectionNetwork.setIouThreshold(.5)
detectionNetwork.setBlobPath("catdog320.blob")
detectionNetwork.setNumInferenceThreads(2)
detectionNetwork.input.setBlocking(False)

# Linking
camRgb.preview.link(detectionNetwork.input)
if syncNN:
    detectionNetwork.passthrough.link(xoutRgb.input)
else:
    camRgb.preview.link(xoutRgb.input)

detectionNetwork.out.link(nnOut.input)

# Connect to device and start pipeline
with dai.Device(pipeline) as device:

    # Output queues will be used to get the rgb frames and nn data from the outputs defined above
    qRgb = device.getOutputQueue(name="rgb", maxSize=4, blocking=False)
    qDet = device.getOutputQueue(name="nn", maxSize=4, blocking=False)

    frame = None
    detections = []
    startTime = time.monotonic()
    counter = 0
    color2 = (255, 255, 255)

    # nn data, being the bounding box locations, are in <0..1> range - they need to be normalized with frame width/height
    def frameNorm(frame, bbox):
        normVals = np.full(len(bbox), frame.shape[0])
        normVals[::2] = frame.shape[1]
        return (np.clip(np.array(bbox), 0, 1) * normVals).astype(int)

    def displayFrame(name, frame):
        color = (255, 0, 0)
        for detection in detections:
            bbox = frameNorm(frame, (detection.xmin, detection.ymin, detection.xmax, detection.ymax))
            cv2.putText(frame, labelMap[detection.label], (bbox[0] + 10, bbox[1] + 20), cv2.FONT_HERSHEY_TRIPLEX, 0.5, 255)
            cv2.putText(frame, f"{int(detection.confidence * 100)}%", (bbox[0] + 10, bbox[1] + 40), cv2.FONT_HERSHEY_TRIPLEX, 0.5, 255)
            cv2.rectangle(frame, (bbox[0], bbox[1]), (bbox[2], bbox[3]), color, 2)
        # Show the frame
        cv2.imshow(name, frame)

    while True:
        if syncNN:
            inRgb = qRgb.get()
            inDet = qDet.get()
        else:
            inRgb = qRgb.tryGet()
            inDet = qDet.tryGet()

        if inRgb is not None:
            frame = inRgb.getCvFrame()
            cv2.putText(frame, "NN fps: {:.2f}".format(counter / (time.monotonic() - startTime)),
                        (2, frame.shape[0] - 4), cv2.FONT_HERSHEY_TRIPLEX, 0.4, color2)

        if inDet is not None:
            detections = inDet.detections
            counter += 1

        if frame is not None:
            displayFrame("rgb", frame)

        if cv2.waitKey(1) == ord('q'):
            break

Does this work for you? Cause it doesn't for me and I think the model is not created correctly.

Thanks ,
Jaka

Hi @li_you_chen
I tested it on images of cats and dogs, with the swapped model inside this example: https://docs.luxonis.com/projects/api/en/latest/samples/Yolo/tiny_yolo/

Could you post the output of detections:

inDet = qDet.tryGet()

if inDet is not None:
            detections = inDet.detections
            print(detections)

You can't interchange Yolo blobs and Mbnet, they have completely different decoding functions.

Thanks,
Jaka

Hi @li_you_chen
I tested with the model and I am getting no detections (to pictures of cats and dogs). I looks like detection problem.

Hi @li_you_chen
Hmm, could you send the model over, along with MRE so we can check please?
Try changing the IOU threshold and the confidence to see if you get any improvements in detections.
Also important make sure you update to the latest depthai version (2.25.1).

Thanks,
Jaka

Hi @li_you_chen
If you train with yolov8, it won't work on mobilenetSSD. Here should be a more up to date yolov8 network notebook which you can use directly.

Thanks,
Jaka

Hi @li_you_chen
I would suggest trying your model on https://docs.luxonis.com/projects/api/en/latest/samples/Yolo/tiny_yolo/. Check if you get proper detections (bounding boxes are correct - IOU and confidence are correct). The tracking should then work as it only uses the detection results.

Thanks,
Jaka

Hi @li_you_chen
You can use ImageManip node to perform a resize. Also, you might have problems with Yolo model expecting a 3 channel input. You can manually convert the GRAY8 image frame into BGR888p inside the same manip node as well.

Similar example: https://docs.luxonis.com/projects/api/en/latest/samples/mixed/rgb_encoding_mono_mobilenet/

Thanks,
Jaka