For my latest deployment on Axelera AI Metis, I’ve been exploring geometric line features for motion estimation, localization and 3D mapping in robotics.
The project, lines-metis, runs DeepLSD inference on Metis with a C++ pipeline that extracts line segments from camera images.
Door frames, wall boundaries and structural edges can provide useful geometry where distinctive point features are sparse. Matched across views and combined with keypoints, depth or IMU measurements, these observations can support visual odometry, SLAM and 3D reconstruction.
The INT8 network runs on one Metis AIPU core, with preprocessing, field decoding and segment extraction on the host CPU. On a 900-frame Freiburg 2 excerpt, the pipeline averaged 27.8 ms/frame, or 35.9 processing FPS, using 200 calibration frames. These timings exclude video I/O and visualization.
The detection pipeline is working. The next useful question is how much these features contribute to localization accuracy when integrated into a robotics estimator.
Code: https://github.com/sanket-pixel/lines-metis
Model: DeepLSD, Pautrat et al., CVPR 2023
Footage: TUM RGB-D, Sturm et al., CC BY 4.0
