Robotour 2025

robotour-2025

depth refinement: input topics

jazzy-v1

Depth Refinement

• Architecture: LinkNet with Mobilenet_v2 and ResNet18 encoders from SMP
• Input: Left Gray Image + Original Disparity
• Output: Refined Disparity
• Training Labels (Disparity): distillation from DEFOM-Stereo
• Data: subtdata:/data/disparity-refinement
• CPU Optimization: with torch.jit.trace, docs
• ROS2: node tested with Luxonis OAK camera (long-range)