Reproduce “Self-Supervised Scale Recovery for Monocular Depth and Egomotion Estimation” on OpenResearch

About this paper

Brandon Wagstaff, Jonathan Kelly

Wagstaff and Kelly introduce a self-supervised monocular learning method for producing metrically-scaled depth and egomotion estimates by integrating a scale recovery loss during training. This approach enables accurate predictions without ground truth or stereo at test time, and allows for effective online retraining, achieving competitive visual odometry accuracy compared to state-of-the-art methods on KITTI and Oxford RobotCar datasets.

Key insights

• Injecting metric scale information directly into the self-supervised training process through specific loss functions is more effective and robust than applying scale correction post-hoc at test time.
• By learning to segment the ground plane and relating estimated camera height to a known height *during training*, the network implicitly learns to produce metrically scaled depth and egomotion outputs without needing external metric references during inference.
• Online retraining, facilitated by this self-supervised metric scale recovery, allows pre-trained models to rapidly adapt to new, out-of-distribution environments, enhancing long-term autonomy.

Summary by alphaXiv.