No detection with webcam mode

[tonginleong]

I tried the normal and v4-tiny model with demo_webcam.py
Everything is at default but no detection can be seen...not a single one, and of coz I tried different angles and stuff but no avail.
I am also a follower of yolov4 btw

[souravrs999]

Try this piece of code

import cv2
import argparse

from yolo import YOLO

arg = argparse.ArgumentParser()
arg.add_argument("-d", "--device", default=0, help="Device to use")
arg.add_argument("-s", "--size", default=416, help="Yolo input size")
arg.add_argument("-c", "--confidence", default=0.4, help="Confidence for yolo")
args = arg.parse_args()

yolo = YOLO(
    "models/cross-hands-yolov4-tiny.cfg",
    "models/cross-hands-yolov4-tiny.weights",
    "hand",
)

yolo.size = int(args.size)
yolo.confidence = float(args.confidence)

print("starting webcam.")
cv2.namedWindow("preview")
cap = cv2.VideoCapture(args.device)

while True:
    if cap.isOpened():
        ret, frame = cap.read()

        if not ret:
            break

        width, height, inference_time, results = yolo.inference(frame)
        cv2.putText(
            frame,
            f"{round(1/inference_time,2)}) FPS",
            (15, 15),
            cv2.FONT_HERSHEY_SIMPLEX,
            0.5,
            (0, 255, 255),
            2,
        )

        for detection in results:
            id, name, confidence, x, y, w, h = detection

            cx = x + (w / 2)
            cy = y + (h / 2)

            color = (0, 255, 255)
            cv2.rectangle(frame, (x, y), (x + w, y + h), color, 2)
            text = "%s (%s)" % (name, round(confidence, 2))
            cv2.putText(
                frame, text, (x, y - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 2
            )

        cv2.imshow("preview", frame)
        key = cv2.waitKey(1)
        if key == 27:
            break

cv2.destroyAllWindows()
cap.release()

[cansik]

@tonginleong Yes you are right, by mistake I merged a PR #10 which breaks the detection.

[cansik]

Fix is uploaded, would be great if you could give me a quick feedback.

[souravrs999]

glad i could help

[cansik]

@souravrs999 The code you posted did not fix the issue, I had to replace the hand-counting mechanism (a939fbe).

[souravrs999]

@souravrs999 The code you posted did not fix the issue, I had to replace the hand-counting mechanism (a939fbe).

It does work I removed the whole handcounting logic which was causing the issue

[tonginleong]

yep, the code is working fine now with the cross-hands.weights and its cfg, thank you.
I dont seem to be able to get it work with v4-tiny, but that is not much of a concern for me.

[thunder95]

I met the same problem, it didnt work with v4-tiny. the results length is always zero. How could I fix this problem?

[cansik]

Without an example image I am not able to help you. The v4-tiny seems to work on my machine. Have you tried lowering the confidence threshold?

python demo_webcam.py -n v4-tiny -c 0.1

But what I can say is that the v4-tiny model does not work as good as the other models on webcam images if you are very close to the webcam. The network did maybe overfit a bit while training.