XPENG RoboticsSep 2021 – Nov 2021

Indoor Semantic Segmentation

An indoor semantic segmentation project comparing transformer-based and CNN-based segmentation models on custom robotics data.

Project Snapshot

Signal Details
Timeline Sep 2021 to Nov 2021
Organization XPENG Robotics
Focus Semantic Segmentation for indoor environments
Stack DDRNet

Model Comparison

Model Family Example Evaluation Focus
Transformer-based SegFormer Accuracy with mIoU
CNN-based DDRNet Real-time speed with FPS

Contributions

  • Selected DDRNet for indoor custom-data segmentation.
  • Used 2D semantic segmentation to handle transparent or glass-like doors and walls where depth estimation is limited.
  • Projected segmented image regions back into point clouds for more accurate semantic 3D perception.