Dataset ingestion and VIO replay

RustRobotics reads the standard extracted layouts of the EuRoC MAV datasets and the KITTI odometry benchmark. Dataset archives are not redistributed by this repository.

EuRoC

Pass either the sequence directory or its mav0 directory to EurocDataset::load. The loader reads:

MH_01_easy/
└── mav0/
    ├── cam0/
    │   ├── data.csv
    │   ├── data/*.png
    │   └── sensor.yaml
    ├── imu0/
    │   ├── data.csv
    │   └── sensor.yaml
    └── state_groundtruth_estimate0/data.csv  # optional

It validates increasing timestamps, parses camera intrinsics, resolution and the camera/IMU T_BS transforms, and exposes efficient IMU interval slices. IMU samples are transformed into the body frame before preintegration, including centripetal and tangential acceleration caused by a non-zero sensor lever arm. A missing legacy imu0/sensor.yaml falls back to an identity transform. Images remain paths so applications can choose their own decoder. RustRobotics includes a PNG-based sparse frontend CLI:

cargo run --release -p rust_robotics \
  --example generate_euroc_feature_tracks \
  --no-default-features --features slam -- \
  /datasets/EuRoC/MH_01_easy

The frontend distributes Shi-Tomasi corners, tracks them with pyramidal Lucas-Kanade optical flow, applies a forward/backward consistency check, and triangulates persistent tracks from the metric IMU-predicted camera trajectory. Only the first ground-truth state initializes pose, velocity and IMU biases. Existing sidecars are never replaced unless --force is passed. Use --output DIR to write elsewhere and --max-features N to adjust the per-frame cap.

The offline VIO example consumes generated or externally extracted tracks from this optional sidecar:

mav0/rust_robotics/
├── landmarks.csv     # landmark_id,x,y,z
└── observations.csv  # timestamp_ns,landmark_id,u,v

Landmark IDs must be contiguous and zero-based. Observation timestamps must match cam0/data.csv. A different frontend can export the same interchange format without coupling the SLAM crate itself to an image library.

# Checked-in miniature replay
cargo run -p rust_robotics --example headless_euroc_vio \
  --no-default-features --features slam

# Extracted sequence after running generate_euroc_feature_tracks
cargo run --release -p rust_robotics --example headless_euroc_vio \
  --no-default-features --features slam -- /datasets/EuRoC/MH_01_easy

The replay uses ground truth only for the initial state and acceptance report. Later ground-truth states are not optimizer inputs:

EuRoC IMU + T_BS ──> lever-arm correction ──> bias-aware preintegration
feature tracks ─────────────────────────────> camera/landmark BA (Schur)
BA poses + IMU factors ─────────────────────> navigation-state/bias refinement
IMU relative edges + visual closure ────────> block-sparse SE(3) pose graph

KITTI odometry

KittiOdometryDataset::load(root, "00") reads the official layout:

dataset/
├── poses/00.txt
└── sequences/00/
    ├── calib.txt
    ├── times.txt
    ├── image_0/ ... image_3/
    └── velodyne/*.bin

The API exposes image/LiDAR paths, 3×4 camera projections, Velodyne-to-camera calibration, timestamps and optional ground-truth poses. read_velodyne decodes the official little-endian (x, y, z, reflectance) float32 tuples.

Checked-in fixtures contain only original synthetic numeric rows in the official layouts; they do not copy dataset imagery or measurements.