Computer vision on a competition robot
· 1 min read
On a competition robot, a model that is 2% more accurate but twice as slow is usually the wrong choice. Latency is part of correctness.
- Use classic OpenCV for simple, well-lit cues; save YOLO for what really needs learning.
- Test under the actual arena lighting as early as possible.
- Keep the robot logic readable — 4,000 lines of ROS code gets hard to change quickly.