OnePole: One-shot recognition of Pole-like objects in Lidar point clouds

Efficient management of Pole-like objects (PLOs) is important in urban environments. Deep learning methods offer potential improvements but depend extensively on large training datasets for high accuracy. This paper presents a novel one-shot learning framework for PLO recognition. Our method requires only a single support sample from each category for effective classification, capitalizing on the prior knowledge that PLOs have very similar features. At the core of our approach is OnePole, a non-parametric framework designed to distinguish different PLOs from other road assets while avoiding the large training datasets typically needed by deep learning models. Our experimental results demonstrate that the proposed method achieves an accuracy of 72.1% as measured by Intersection over Union (IoU) on PLOs, a performance comparable to that of state-of-the-art fully supervised deep learning methods. Our method achieves this with only a single labelled object per category, substantially enhancing the efficiency of detecting PLOs in point cloud datasets. This not only illustrates the effectiveness of our approach but also represents a significant advancement in reducing the reliance on extensive labelled data, which is a common challenge in fully supervised methods.