XPENG RoboticsFeb 2022 – May 2022

Uncertainty Modeling for Stereo Depth Estimation

A stereo depth estimation project focused on modeling both aleatoric and epistemic uncertainty for safer real-time depth prediction.

Project Snapshot

Signal Details
Timeline Feb 2022 to May 2022
Organization XPENG Robotics
Focus Aleatoric and Epistemic Uncertainty Modelling in Stereo Depth Estimation
Stack Custom implementation and project-specific tooling

Approach

  • Trained a Deep Learning Cost Volume based Stereo Depth Estimation Algorithms on custom Sparse and Dense datasets for performance comparison
  • Aleatoric Uncertainty Modelling by learning Variance by adding a small Variance Network to current Depth Estimation Algorithm and training with NLL (Negative Likelihood) Loss
  • Initially learning Epistemic Uncertainty using MC-Dropout Method (Requires multiple Forward passes)
  • Explored Sampling-free Epistemic Uncertainty Estimation using Evidential Deep Learning
  • To Deploy sampling free Uncertainty Estimation to Filter Uncertain Depth Prediction on a real time system

Goal

Deploy sampling-free uncertainty estimation to filter unreliable depth predictions in a real-time robotics perception system.