Stereo Perception & Semantic Segmentation

A computer vision research system combining semantic segmentation and stereo depth estimation for robot-environment understanding.

Dense spatial perception under strict size, weight, and power constraints required both an efficient model and a scalable data pipeline.

Built the PyTorch model, automated annotation for 50,000 stereo images with YOLOv8, and integrated CREStereo depth estimation.

Stereo input passes through segmentation and disparity estimation to produce a denser spatial representation for downstream robotic perception.

1. 01

   Automated annotation with YOLOv8

   Instead of

   Manual labeling of the dataset

   Why

   Hand-labeling 50,000 stereo pairs was not feasible on the project timeline. Automated annotation introduced label noise, which was the accepted cost — the dataset scale mattered more to model performance than per-label precision.

2. 02

   CREStereo for dense disparity

   Instead of

   Classical block matching (SGBM) or monocular depth estimation

   Why

   Block matching degrades badly on the textureless surfaces common indoors, and monocular estimation gives relative rather than metric depth. CREStereo produced dense metric disparity at the cost of significantly more compute.

3. 03

   Tuned for 11 FPS under SWaP constraints

   Instead of

   Maximizing accuracy offline

   Why

   The pipeline had to run on power- and weight-limited onboard hardware. Model size and input resolution were traded down to hold a usable real-time rate, since perception that arrives late is not useful to a robot in motion.

• Python
• PyTorch
• OpenCV
• YOLOv8
• CREStereo