Low-Compute Trash Segmentation Using Pretrained Vision Transformers Without Fine-Tuning

With more and more people in our world, there is unfortunately more and more garbage, especially in the environment. Collecting it can be tedious and requires valuable human labor. On the other hand, there are several projects for autonomous robots to take over this task. However, visual perception of garbage is still an open challenge due to its variance and diversity of objects. We propose a computationally inexpensive perception architecture that uses the attention-based skills of pretrained vision transformers to locate and coarsely segment trash in images. No special fine-tuning of the model for trash objects is required. In addition, we incorporate a distraction branch into our architecture based on LRP, which provides the ability to divert attention away from common objects, such as small animals, that were part of the initial general pretraining. Our experiments demonstrate the advantages of our approach when new classes are introduced during testing. Furthermore, the modular architecture enables future developments towards computationally efficient trash segmentation for mobile robots.