Unsupervised Learning for Structure from Motion

Sammanfattning: Perception of depth, ego-motion and robust keypoints is critical for SLAM andstructure from motion applications. Neural networks have achieved great perfor-mance in perception tasks in recent years. But collecting labeled data for super-vised training is labor intensive and costly. This thesis explores recent methodsin unsupervised training of neural networks that can predict depth, ego-motion,keypoints and do geometric consensus maximization. The benefit of unsuper-vised training is that the networks can learn from raw data collected from thecamera sensor, instead of labeled data. The thesis focuses on training on imagesfrom a monocular camera, where no stereo or LIDAR data is available. The exper-iments compare different techniques for depth and ego-motion prediction fromprevious research, and shows how the techniques can be combined successfully.A keypoint prediction network is evaluated and its performance is comparedwith the ORB detector provided by OpenCV. A geometric consensus network isalso implemented and its performance is compared with the RANSAC algorithmin OpenCV. The consensus maximization network is trained on the output of thekeypoint prediction network. For future work it is suggested that all networkscould be combined and trained jointly to reach a better overall performance. Theresults show (1) which techniques in unsupervised depth prediction are most ef-fective, (2) that the keypoint predicting network outperformed the ORB detector,and (3) that the consensus maximization network was able to classify outlierswith comparable performance to the RANSAC algorithm of OpenCV.