Abstract:
Due to its key role in many real world applications such as autonomous driving, semantic segmentation often has to be applied on edge devices with limited computational r...Show MoreMetadata
Abstract:
Due to its key role in many real world applications such as autonomous driving, semantic segmentation often has to be applied on edge devices with limited computational resources. At runtime, the requirements for the segmentation system can change significantly depending on the given circumstances. For example, in some cases the focus is on a fast segmentation of certain particularly important classes, in other cases it is more relevant to distinguish a large number of classes. Besides many works aiming at a resource efficient but still accurate semantic segmentation, the possibility to adapt the segmentation model to specific circumstances is missing. In this work, we present a novel approach for flexible semantic segmentation that builds on recent developments in segmentation foundation models and prompt tuning. The approach offers the possibility to trade off inference time and accuracy, to flexibly select the classes to be segmented, and to add new target classes to an existing system just by transferring new prompts. Evaluation on an edge device shows that the inference time can be significantly reduced with fewer classes and smaller prompts, and that accuracy increases with larger prompts at the expense of a longer inference time.
Date of Conference: 19-22 May 2024
Date Added to IEEE Xplore: 02 July 2024
ISBN Information: