Balancing Accuracy and Efficiency for River Plastic Monitoring

The availability and use of drones, fixed cameras, and smartphones allows the advancement in close range optical remote sensing due to its operability and viability for large-scale implementation. Specifically, RGB cameras are particularly advantageous for low-cost close-range sensing, offering higher spatial resolution at maximum two orders of magnitude compared to satellite sensors. This characteristic proves invaluable in environmental monitoring applications, such as tracking plastic movement in riverine environments. However, one of the pressing issues in object detection is to find a balance between accuracy and computational cost even with single stage object detection algorithms such as You Only Look Once (YOLO). Considering the environmental variabilities such as weather and different lighting conditions, here we attempt to demonstrate uncomplicated data augmentation techniques for both train and test images with primary focus on lighting adjustment. The goal is to determine which approach will increase performance with minimum cost.