Efficient Deep Learning Models for Litter Detection in the Wild

Littering presents a substantial environmental hazard and impacts our well-being. The importance of automatic litter detection lies in its ability to identify waste in the environment, thereby enhancing the efficiency of subsequent waste management operations. In order to achieve a comprehensive and detailed survey of an area for litter detection, one of the most effective approaches is to utilize the collective efforts of citizen science. In this work we assess the performance of the most efficient object detection methods aiming their use in the type of devices typically employed in citizen science activities, e.g. smartphones with low processing capabilities. Experiments on the Trash Annotations in COntext (TACO) dataset show that by exploiting our training procedure, the efficient models that we tested are able to surpass the performance reached by larger models in the state of the art. Moreover, experiments show the among the efficient object detectors tested, the small model variants offer the best trade off between model size and litter detection performance.