Leveraging EfficientNetV2 for Litter Detection

Litter pollution is a significant environmental challenge globally, demanding advanced methods for detection and management. Recent advancements in computer vision offer promising alternatives to traditional approaches, yet they often rely on narrowly focused datasets with a lack of segmentation data, which is required for applications such as robotic grasping. This research paper addresses these limitations while utilising the TACO (Trash Annotations in Context) dataset, a small but growing dataset, created to encompass diverse environmental contexts and detailed annotation. EfficientNetV2, a state-of-the-art convolutional neural network architecture, is leveraged for both image classification and object detection/instance segmentation models. The image classification model achieves an F1-score of 0.854 in categorizing an image containing ‘Litter' or ‘No Litter'. For object detection and instance segmentation, EfficientNetV2 is integrated with the Mask R-CNN architecture implemented in Detectron2. This hybrid model achieves a box and segmentation mAP of 0.65 and 0.52 respectively.