Few-shot Traversability Segmentation of Indoor Robotic Navigation with Contrastive Logits Align

Qiyuan An, Christos Sevastopoulos, Farnaz Farahanipad, Fillia Makedon

Abstract: Indoor traversability segmentation is of paramount importance for mobile robotic navigation and path planning as it identifies the free-space where the robot can safely traverse. This paper proposes a Few-Shot Learning (FSL) approach to meta-learn an existing pretrained segmentation model for an indoor traversability segmentation task. Our goal is to meta-learn a segmentation model in a sense that it can adapt to a new unseen class given limited annotated examples. Directly applying FSL on this task often results in non-ideal segments across various scenarios, due to the limited samples of the new class. To deal with this issue, we propose the Contrastive Logits Align (CLA) method that allows the segmentation model to infer free-space not only from similar positive segments, but also from the opposite class of obstacles. Inspired by the spirit of contrastive learning, our proposed method first adapts the classifier of the segmentation model on the support set and extract prototypes for both obstacles and free-space on the logits layer. Additionally, CLA also extracts negative prototypical features of obstacles and then infers free-space by discriminating the difference between obstacles and free-space. We conduct extensive experiments on an indoor traversability dataset collected from various university campus buildings and show satisfactory performance of the effectiveness and generalization ability of our proposed method to unseen cases.

External IDs:dblp:conf/case/AnSFM24