Bounding boxes here appear slightly behind due to network latency, by the time the frontend receives them, the video has progressed a ~1-3 frames
I wanted to learn more about training object detection models for applications in agriculture. I couldn't really find/collect any data of a tractor going over a field so I stitched together images that closely resmbled this, this actually helped solve the issue of training in different lighting environments and color attributes of the soil like hue, saturation, and intensity.
In practice something like the Carbon Robotics LaserWeeder would normalize lighting conditions as much as possible by having LEDs near the optical sensors, also allowing for autonomous use throughout the night.
This was an interesting weekend project and as you can see the performance is not fantastic (limited on data—about 10k images total for training and eval) but was able to test and validate some of my ideas of training models that face occlusion and more environmental vairance.