rice

https://aidea-web.tw/topic/9c88c428-0aa7-480b-85e0-2d8fb2fcf3fc?focus=team

Dataset format

• Original:
  • Some are 2000 x 3000, some are 1728 x 2304
  • Original csv data are labeled in x, y format

Dataset preparation

- split train/val manually
- `yolov5/myutils/image_cutter.py`
- `yolov5/myutils/csv2yolo.py`

Try

• TODO

  • Add a green-only-channel(or only)
  • Try custom mean/variance for normalization
  • Model with big kernel(100 x 100) to find the patterns of the grid.
  • Cut without resizing
  • Larger bbox for d-img
  • Decrease weight of cls loss in total loss
  • yolo 1280
  • Train all
  • Use DBSCAN to filter extreme points

• Tried and beneficial

  • Merge the edge
    • Can be further improved by sliding window + nms
  • Decrease IOU threshold
  • Decrease conf threshold
  • Add hsv augmentation to detect rice in very bright/dark situations
    • Does not help since bright/dark situations won't be labeled in the ground truth

• Tried but doesn't help

  • Use different detectors accordingly
  • Stride training

Learn

• 不要一開始就找完美解答
• 大model != 高準確度