Excavate Condition-invariant Space by Intrinsic Encoder.

Jian Xu,Chunheng Wang,Cunzhao Shi,Baihua Xiao

DOI: https://doi.org/10.48550/arxiv.1806.11306

2018-01-01

Abstract:As the human, we can recognize the places across a wide range of changingenvironmental conditions such as those caused by weathers, seasons, andday-night cycles. We excavate and memorize the stable semantic structure ofdifferent places and scenes. For example, we can recognize tree whether thebare tree in winter or lush tree in summer. Therefore, the intrinsic featuresthat are corresponding to specific semantic contents and condition-invariant ofappearance changes can be employed to improve the performance of long-termplace recognition significantly. In this paper, we propose a novel intrinsic encoder that excavates thecondition-invariant latent space of different places under drastic appearancechanges. Our method excavates the space of intrinsic structure and semanticinformation by proposed self-supervised encoder loss. Different from previouslearning based place recognition methods that need paired training data of eachplace with appearance changes, we employ the weakly-supervised strategy toutilize unpaired set-based training data of different environmental conditions. We conduct comprehensive experiments and show that our semi-supervisedintrinsic encoder achieves remarkable performance for place recognition underdrastic appearance changes. The proposed intrinsic encoder outperforms thestate-of-the-art image-level place recognition methods on standard benchmarkNordland.