---
license: cc-by-4.0
pretty_name: Minimal Texture Dataset for RGB-D SLAM
tags:
  - slam
  - rgbd
  - visual-odometry
  - robotics
  - realsense
---

# Minimal Texture Dataset for RGB-D SLAM

RGB-D sequences recorded with an Intel RealSense D435i over low-texture conceptual-geometry and
sand scenes at DLR (German Aerospace Center), with millimeter-accurate ground truth from a Vicon
motion-capture system. The dataset was introduced for the evaluation of
[SID-SLAM: Semi-Direct Information-Driven RGB-D SLAM](https://ieeexplore.ieee.org/document/10058038)
(IEEE RA-L, 2023).

This repository is the authoritative re-upload of the dataset previously published on
[Zenodo (10.5281/zenodo.10453700)](https://zenodo.org/records/10453700), regenerated from the
original recordings. It extends the published set with the full first recording session and the
previously unpublished easy sequences.

- Homepage: https://www.dlr.de/en/rm/research/publications-and-downloads/datasets/minimal-texture-dataset-for-rgb-d-slam
- Integrated in [VSLAM-LAB](https://github.com/VSLAM-LAB/VSLAM-LAB) as dataset `minimal`

## Contents

Sequences are grouped by recording day:

- `10-08/` — first session, 2021-10-08 (15 sequences: parallel_lines, circles, hexagon, triangle,
  deformation_model, sand, square_with_circles, and loop variants)
- `10-14/` — second session, 2021-10-14 (18 sequences: the 14 published on Zenodo — triangle,
  circle, square, hexagon, dodecagon, lines, the three circle_* loop sequences, sand_rocks — plus
  forward_backward_01, circle_only_01, lru_01 and ardea_01)

Each sequence ships two files:

- `<day>/<name>.bag` — ROS1 bag (bz2-compressed) with the RealSense D435i streams:
  - `/camera/color/image_raw` — rgb8, 1280x720 @ 30 Hz
  - `/camera/aligned_depth_to_color/image_raw` — 16UC1 depth aligned to color, millimeters, 1280x720 @ 30 Hz
  - `/camera/color/camera_info`, `/camera/aligned_depth_to_color/camera_info` — intrinsics
  - `/camera/imu` — ~200 Hz (camera-to-IMU extrinsics not included)
  - `/camera/extrinsics/depth_to_color`
- `<day>/<name>_groundtruth.txt` — Vicon ground-truth trajectory in TUM format
  (`timestamp[s] tx ty tz qx qy qz qw`), evaluation-ready, same clock as the bag.

`T_vicon_camera.txt` (repo root) holds the Vicon-marker to `camera_color_optical_frame` extrinsic.

## Calibration

Color intrinsics (identical for both sessions, zero distortion / plumb_bob, 1280x720):

```
fx = 908.479248  fy = 907.841919
cx = 650.889160  cy = 364.692657
```

The aligned depth stream shares these intrinsics by construction. Depth in meters = value / 1000.

## Calibration recordings & raw metadata (archive)

The original calibration recordings and per-sequence provenance from the 2021 sessions are
archived here as well (not needed to run the sequences, but preserved as the primary copy):

- `calibration/10-08/` — `timestamps_camera_calibration_seq01.bag`,
  `viconMarker_camera_calibration_seq01.bag`, `viconMarker_camera_calibration_seq02.bag`
- `calibration/10-14/` — `deformation_model_calibration.bag`,
  `lukas_pattern_calibration_01.bag`, `photo_calibration.bag`

These are the bz2-compressed ROS1 bags used for the Vicon-marker/camera extrinsic
(`T_vicon_camera.txt`), camera-clock/Vicon timestamp calibration, and photometric/pattern
calibration.

- `metadata/10-08.tar.gz`, `metadata/10-14.tar.gz` — the original per-sequence recording folders
  (environment and git-version snapshots, `params.yaml`, `tf.cfg`, raw Vicon/tf bags, and
  `vicon_direct_features_*.txt`), documenting how each sequence's ground truth was produced.

## Citation

```bibtex
@article{fontan2023sid,
  title={{SID-SLAM}: Semi-Direct Information-Driven {RGB-D SLAM}},
  author={Font{\'a}n, Alejandro and Giubilato, Riccardo and Oliva, Laura and Civera, Javier and Triebel, Rudolph},
  journal={IEEE Robotics and Automation Letters},
  volume={8},
  number={10},
  pages={6387--6394},
  year={2023}
}
```

## License

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)