Explore use of automatic workflows to find bears in DenCam data

[iingram]

Some combination of a fast OD (COCO has bears so many OTS YOLO models have bears) and striding through frames should make it feasible to somewhat rapidly process the videos.

We should always be weighing this against other options like simply watching the video really fast in VLC or using CV-based motion detection and/or background subtraction that will be able to run even faster through the frames i.e. leveraging how little is happening in these views when bears aren't there.

[kgarwoodsdzwa]

using yolo_tracker_save.py from frangle repo, used yolo11n on hornsund 2024-03-22-10h59m36s.mp4 with known polar bear emergence and it was able to successfully save the video containing the polar bear with tracks. it only "detected" it for 15 seconds out of the 5 minutes, despite being in the frame and visible for the whole video. it thought it was either a person or a kite. will likely need a different approach as it was also super slow, could try a different yolo version, this was just the smallest of the newest versions. or use md5 or 6 doing detection of any of the classes on about a frame per second to start

[kgarwoodsdzwa]

tried this tool https://github.com/sheneman/tigervid, using yolov8x, interestingly, it only detected times when the cub was out and ignored the mother in that clip need to look into that more. liked that it allowed you to choose how many seconds to skip before analyzing a frame, and then output a video with a chosen buffer around the detection