---

name: raw-data description: "Load and validate KITTI raw drives with pykitti.raw, including calibration, images, Velodyne scans, timestamps, and OXTS poses." disable-model-invocation: true metadata: disco-role: operating license: MIT

KITTI raw drives

Use this sub-skill when the input is an original KITTI raw date/drive tree and the operation is pykitti.raw. It covers the raw directory layout, calibration, timestamps, OXTS packets/poses, four camera streams, stereo pairs, Velodyne scans, and positional frame subsets.

Route boundaries

• Use odometry for pykitti.odometry, sequences/<sequence>, times.txt, and benchmark pose files.
• Use tracking for tracking data and labels.
• Use the package-level troubleshooting guide for installation, package import, and workflow escalation.
• Use the local OpenCV/stereo guide only after raw images have been loaded as Pillow images.

Prerequisites and safety

The covered release is pykitti==0.3.1; install its declared dependencies (numpy, Pillow, matplotlib, and pandas) in the execution environment. The normal top-level import pykitti also imports tracking.py, which imports cv2 eagerly. Therefore OpenCV is needed for a successful top-level import in this release even though raw.py uses Pillow rather than OpenCV to decode raw images. See troubleshooting.

Do not download or unpack KITTI data from a runtime workflow. Acquire and extract archives separately, then validate the tree in data layout. The constructor requires the three date-level calibration files and oxts/timestamps.txt; it does not validate sensor counts and an empty OXTS data directory can produce an empty oxts list.

Quick start

import pykitti

data = pykitti.raw(
    base_path, "2011_09_26", "0019",
    dataset="sync",       # default
    frames=[0, 5, 9],      # positional indices; default None means all
    imtype="png",         # extension without the leading dot
)

print(len(data), data.timestamps[0])
left = data.get_cam0(0)       # PIL.Image, mode L
left_rgb, right_rgb = data.get_rgb(0)
scan = data.get_velo(0)      # float32, shape (N, 4)

base_path and date are joined, and the drive directory is <date>_drive_<drive>_<dataset>. The constructor eagerly reads calibration, timestamps, and the selected OXTS files; image and Velodyne bytes are lazy. Use a reusable list or materialized range for frames, and validate every stream before pairing it. frames selects positions in independently sorted file lists, not filename IDs. A valid non-contiguous example is [2, 0].

Access and invariants

• len(data) is the number of selected timestamps.
• cam0/cam1 yield fresh Pillow L images; cam2/cam3 yield fresh Pillow RGB images.
• gray and rgb are one-pass zip generators. Indexed get_gray(i) and get_rgb(i) are preferable when a frame is needed repeatedly.
• velo is a fresh generator of float32 arrays shaped (N, 4) with columns [x, y, z, reflectance]; get_velo(i) reads one selected-list position.
• timestamps contains naive datetime.datetime values. KITTI timestamp nanoseconds are truncated to microseconds by this release.
• oxts contains OxtsData(packet, T_w_imu) records. The first pose is origin-normalized but is not necessarily an identity rotation.
• calib is a named tuple with 3x4 projection matrices, 4x4 transforms, 3x3 intrinsics, and b_gray/b_rgb baselines. The tuple structure is fixed, but its contained NumPy arrays remain mutable.
• T_camN_velo maps homogeneous Velodyne points to camera-N coordinates. For projection, compute P_rect_20 @ T_cam2_velo @ p_velo and divide by positive depth only.

Read API reference for exact fields and shapes, workflows for repeatable load recipes, and troubleshooting for recovery actions.

Verification

Run the self-contained fixture before using a real drive:

python scripts/raw_fixture_smoke.py --help
python scripts/raw_fixture_smoke.py

It creates and removes a tiny local raw tree by default, exercises non- contiguous selection, and asserts timestamp/OXTS, calibration, image, stereo, and Velodyne shapes. It performs no network or GUI operations. For real data, compare numeric frame IDs and counts across all four cameras, Velodyne files, OXTS files, and timestamp lines before consuming synchronized pairs.