← all projects

Beach trash from the air

VIT · 2024 · with Meghana V Gaonkar, Leona E Joseph, advisor: Dr. R Balamurugan

  • yolov8
  • uav
  • wise-iou

problem

plastic on shorelines is hard to survey on foot; drones see a lot of coast, but something has to find the debris in the images.

what I built

  • a benchmark of five YOLOv8 variants (n, s, m, l, x) with and without a Wise-IoU loss
  • hyperparameter tuning and custom augmentation on a dataset combining two UAV sources (772 beach images and 40 river images), all labelled as one class, garbage
  • the 40 river images were augmented (blur, rotation, scale) into 120, for 892 images in all (772 + 120), split 636 / 128 / 128 into train, validation and test

results

  • YOLOv8m went from 81.7% to 86.1% mAP@50
  • Wise-IoU hurt the largest models: YOLOv8l and YOLOv8x lost a lot of precision, and the paper suggests it may not scale well with larger models
  • hyperparameter tuning (rotation 45 degrees, flips 0.5 and 0.6, MixUp 0.05) lifted YOLOv8s from 80.8% to 83.5% mAP@50
two aerial photos, one of green water and one of a debris-covered shore, with red boxes and confidence scores around detected trash
yolov8 detections on drone shots of water and shoreline
two precision-recall curves for yolov8m, one at 0.818 mAP@0.5 and one at 0.861 mAP@0.5, the second hugging the top right corner more closely
yolov8m precision-recall curves, standard vs tuned hyperparameters and wiou
open paper

open paper · download

← all projects