How ethz-asl/maplab works
Project type and purpose
• This is maplab 2.0, a modular C++ research framework for visual-inertial multi-session and multi-robot mapping, with both command-line and ROS‐based applications (ROVIOLI, maplab-node, maplab-server-node), algorithm libraries, data structures, back-end services, and visualization.
High-level architectural layers
• Layer 1: Hardware & External Services
• Sensors (camera, IMU, LiDAR) via ROS topics or rosbag
• External “building blocks” (ROVIO, HNSW, PMVS, Docker/ROS master)
• Layer 2: Core Libraries
• aslam_cv2 (camera models, frames, detectors, trackers, geometric vision)
• common modules (vio-common, maplab-common, console-common, message-flow, serialization, resources)
• algorithms (feature-tracking, triangulation, dense reconstruction, IMU integration, loop closure, map optimization/sparsification/anchoring)
• Layer 3: Map Data Structures
• -structure packages: sensors, posegraph, localization-summary, vi-map
• backend: map-resources, map-manager for on-disk caching and I/O
• Layer 4: Applications & Plugins
• console apps (maplab-console, console-plugins) for interactive workflows
• ROS nodes (rovioli, maplab-node, maplab-server-node) for online VIO, mapping, multi-robot coordination
• maplab-launch (preconfigured launch files and sensor calibration data)
• Layer 5: Visualization & Tools
• visualization library (RViz plugins, plotters)
• tools (resource-importer, CSV export, test apps)
• deploy (Dockerfiles, scripts, CI)
Main components and interactions
a. Sensor Input → ROVIO / Synchronizer Flow
• ROS topics or rosbag → imu-camera-synchronizer → ROVIOLI node
• Produces odometry estimates, feature tracks
b. Feature Tracking & Map Building
• feature-tracking pipeline → landmark triangulation → stream-map-builder (online)
c. Loop Closure & Optimization
• loop closure modules detect matches → loop-closure-handler → map-optimization builds optimization problem → solver (Ceres)
d. Map Storage & Management
• vi-map data structure stored via map-resources → map-manager for caching and retrieval
e. Console & Plugins
• maplab-console orchestrates operations via message-flow bus → loads map, invokes plugins (dense-recon, sparsification, optimization) → writes out updated map
f. Visualization
• visualization sink publishes markers to RViz → user inspects trajectories, landmarks, map quality
Architectural patterns and principles
• Layered modularity: clear separation between perception, algorithms, data, back-end, and UI
• Plugin system: console plugins and ROS nodes loaded dynamically via shared libraries and gflags
• Message bus: common/message-flow for in-process pub/sub abstraction
• Protobuf & YAML: for serialization (maplab-common, aslam-serialization) and configuration
• CMake/Catkin: each module uses its own CMakeLists and package.xml to express dependencies
Detailed capture of essential elements
• map-structure hierarchy (sensors → posegraph → vi-map → summary) feeding into backend
• console-common/message-flow as a bus backbone connecting console and plugins
• With technology labels: C++, Protobuf, gflags, glog, OpenMP, ROS Melodic, Docker