Train YOLOv8n and performance on down-cam to verify performance #57
Description
🎯 Objective
Train a Robosub dataset captured from competition on a YOLOv8n model and infer on the down-cam. This should be done for both the real-life dataset captured from competition, but also for a synthetic dataset captured from a simulation.
📖 Description
Train the YOLOv8n model, then perform inference and verify the performance. Document the findings both using diagrams (loss functions for example), and your findings. Re-train the model and document what changes you have made to make sure that this worked.
✅ Acceptance Criteria
-
YOLOv8n.pt file trained - Verify both the mAP50 and IOU values within simulation
- Verify in real-life camera during a pooltest
🛠️ Implementation
Approach: Start by just training a model and prototyping a really lightweight inference node that would downsample the downcam resolution to the YOLO model resolution. Add some inference on this and document.
Dependencies:
🔗 Links
Related: [#issue], [#issue]
Blocks: [#issue]
Blocked by: [#issue]