Terrain Traversability Prediction through Self-Supervised Learning and Unsupervised Domain Adaptation on Synthetic Data
Giuseppe Vecchio, Simone Palazzo, Dario C. Guastella, Daniela Giordano, Giovanni Muscato and Concetto Spampinato
This is the official repo for paper "Terrain Traversability Prediction through Self-Supervised Learning and Unsupervised Domain Adaptation on Synthetic Data".
Dataset, code, and additional resources will be made available here.
Terrain traversability estimation is a fundamental task for supporting robot navigation on uneven surfaces. Recent approaches based on deep models for predicting traversability from RGB images have shown promising results; however, supervised training of such models requires manual annotation of a large number of images. In this paper, we present a learning approach for traversability estimation on unlabeled videos by combining dataset synthesis (with computer-generated annotations), self-supervision and unsupervised domain adaptation. Our method poses traversability estimation as a vector regression task over vertical portions of the observed scene. The regression model is first pre-trained through self-supervision to reduce the distribution shift between synthetic and real data and encourage shared feature learning. Then, the model is trained in a supervised way on synthetic videos, while employing an unsupervised domain adaptation loss, by means of gradient reversal, to improve its generalization capabilities on real scenes. Experimental on-field results show that our approach is able to achieve an estimation accuracy on par with standard supervised training, without the need to carry out manual annotation.
The dataset employed in the experiments reported in the paper is available at this link (1.42 GB zip file).