How simondemoreau/doc-depth works
Project type and purpose
• DOC-Depth is a multi-stage data-processing pipeline (not a web app or microservice). It ingests raw sensor data (ROS “.bag” files), performs camera–LiDAR calibration, pre-processes the data with SLAM & classification, and finally applies a custom rendering pipeline (with GLSL shaders) to produce dense depth ground truth on KITTI scenes.
File structure
• Top-level folders:
• Calibration: Python tools to extract frames, detect chessboards, solve LiDAR–camera extrinsics.
• Preprocessing: Python pipeline for pose extraction (SLAM), coordinate transforms, reconstruction classification via DOC.
• Rendering: Python application plus GLSL shader files to composite LiDAR classifications into dense depth maps.
• Each stage has its own “src” subfolder with modular code (e.g., camera.py, lidar.py, transform.py).
• There are no heavy build scripts or containers—this is a script-driven pipeline reliant on ROS and Python packages.
README for architecture hints
• Three sequential stages (Calibration → Preprocessing → Rendering).
• External dependencies: ROS (for recording), a SLAM system (e.g., ORB-SLAM2), and the DOC algorithm for classification.
• Input/Output per stage:
• Calibration inputs: raw .bag files → outputs: aligned point clouds + camera intrinsics/extrinsics.
• Preprocessing inputs: calibrated data → outputs: dense, classified 3D reconstruction + trajectories.
• Rendering inputs: classified point clouds + poses → outputs: per-pixel depth images.
a. Main components
• Data Ingestion (ROS Bag Recorder)
• Calibration Module
• Preprocessing Module
• Rendering Module
• External Services: SLAM engine, DOC classifier
• Data Store / File System (for intermediate artifacts)
b. Relationships & data flow
1. Raw sensors → ROS Bag
2. ROS Bag → Calibration
3. Calibration → calibrated point clouds + camera parameters → Preprocessing
4. Preprocessing ↔ SLAM engine & DOC → reconstructed, classified 3D model + trajectories → Rendering
5. Rendering + shaders → dense depth maps
c. Architectural patterns
• Linear, modular pipeline.
• Each folder is a self-contained stage with clear inputs/outputs.
• Pluggable external engines (SLAM, classification).
d. Key technologies
• Python scripts (numpy, OpenCV, PCL bindings)
• ROS (.bag API)
• SLAM library (external)
• DOC classification code
• OpenGL/GLSL shaders (fragment and vertex programs)