Bug - Inference on videos isn't working as expected?

[utility-aagrawal]

Hi All / @fcakyon ,

I would like to bring something to your attention which I discovered during one of my experiments. I am working on detection of small objects. I tested SAHI with YOLOv8 on some of my test videos and noticed that results/detection were different, compared to YOLOv8 without SAHI results. I used get_prediction() first to ensure that results were consistent with my just YOLOv8 results but they were not.
To debug the issue, I extracted frames from a test video and used get_prediction() on these frames and results matched with just YOLOv8 results. I realized something was wrong with the reading videos part get_video_reader(). get_video_reader() uses opencv to read images/frames and they are in BGR format while get_prediction() expects input image in RGB format. I made this change and everything worked as expected.

Please let me know your opinion on this. Let me know if you have any questions. If you'd like me to open a PR to fix these issues, let me know. Thanks!