YOLOv7 Demo flips the Red and Blue channels on inference

[elijah-utopia]

The Video class that reads from a variety of image sources and returns "regular OpenCV frames" seems to return frames in the BGR format when the model autoshape function interprets at RGB frames.

It seems the expected behavior for python autoshape is that cv images are meant to have their channels already flipped
https://github.com/ultralytics/yolov5/blob/30e4c4f09297b67afedf8b2bcd851833ddc9dead/models/common.py#L243-L252

This is a model trained to recognize red rovers as targets and blue rovers as allies

This is what happens with the yolov7 norfair tracker demo

This is what happens with the demo modified with reversing the channels

        yolo_detections = model(
            frame[:,:,::-1], # convert from cv format BGR to np format RGB
            conf_threshold=args.conf_threshold,
            iou_threshold=args.iou_threshold,
            image_size=args.img_size,
            classes=args.classes,
        )

Environment :

• OS: Ubuntu 20.04
• Python version: 3.8.10
• Norfair version: commit 009a1b1 (master branch as of July 5, 2024)