Feat: Benchmark CV strategies

[kimichenn]

Currently, YOLO v11 (both fine tuned and non-fine tuned) don't work that well. We want to test out different SOTA models, including:

• Owlv2 (has already shown good results from initial testing)
• Grounding DINO
• DINO v3
• mmdet

We want to look into finetuning these models on humans / mannequins from different angles surrounded by bushes, as well as tents. Don't run your finetuning locally if it's going to take over a day, use Datahub or just ask me to run it on a hpc.

Try to look for datasets online that has tents and mannequins in different positions from an aerial view. If not, use not stolen. We should contain ones that has noise (trees, bushes, cars, etc) around the mannequin, as well as just the mannequins themselves. The mannequins may or may not be clothed, so we should consider both types.

[kimichenn]

From some research / testing for my lab:

• Grounding Dino's dependency is cancerous. Can't set it up properly.
• DINO v3 is just a foundational model, meaning that it only extract features, but not object detections.
• mmdet is just a glorified wrapper that's outdated and really annoying to work with.

As such, we should stick with Owl v2