Learning to See by Moving

TL;DR

This paper explores using egomotion as a supervisory signal for feature learning in computer vision, demonstrating it can rival traditional label-based methods across various visual tasks.

Contribution

It introduces egomotion as a novel, freely available supervision signal for training visual features, inspired by biological perception during movement.

Findings

Egomotion-based features perform well on scene and object recognition.

Egomotion supervision compares favorably to label-based training.

Features learned from egomotion aid visual odometry and keypoint matching.

Abstract

The dominant paradigm for feature learning in computer vision relies on training neural networks for the task of object recognition using millions of hand labelled images. Is it possible to learn useful features for a diverse set of visual tasks using any other form of supervision? In biology, living organisms developed the ability of visual perception for the purpose of moving and acting in the world. Drawing inspiration from this observation, in this work we investigate if the awareness of egomotion can be used as a supervisory signal for feature learning. As opposed to the knowledge of class labels, information about egomotion is freely available to mobile agents. We show that given the same number of training images, features learnt using egomotion as supervision compare favourably to features learnt using class-label as supervision on visual tasks of scene recognition, object recognition, visual odometry and keypoint matching.