XPENG RoboticsNov 2021 – Feb 2022

Real-time Aleatoric Uncertainty Estimation in Semantic Segmentation

Added real-time aleatoric uncertainty estimation to a semantic segmentation pipeline so uncertain regions can be filtered before downstream use.

Project Snapshot

Signal Details
Timeline Nov 2021 to Feb 2022
Organization XPENG Robotics
Focus Real-time Aleatoric Uncertainty Estimation in Semantic Segmentation
Stack DDRNet, ConfidNet

Approach

  • Added a ConfidNet FCN uncertainty branch to the existing DDRNet segmentation network.
  • Trained the model to estimate segmentation uncertainty alongside class predictions.
  • Used thresholding to filter uncertain segmented areas in real time.