Voxel51/TUM-RGBD

Dataset Card for TUM RGB-D SLAM benchmark

This is a FiftyOne dataset with 47 samples. Each sample is one recorded sequence from the benchmark, stored as a native multimodal MCAP episode.

The source is the TUM RGB-D SLAM Dataset and Benchmark from the Computer Vision Group at the Technical University of Munich. A Microsoft Kinect records color and depth at 640x480 and 30 Hz while an eight-camera motion-capture system tracks it at 100 Hz. The sequences are recorded with three Kinects, which the benchmark names freiburg1, freiburg2 and freiburg3, handheld, on a Pioneer robot, and over scenes built to test structure against texture, moving people and object reconstruction. This conversion carries the 47 sequences of the benchmark's main table, every one with its ground truth.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = fouh.load_from_hub(
    "Voxel51/TUM-RGBD",
    name="TUM-RGBD",
    persistent=True,
)

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

47 sequences, 9 recorded with freiburg1, 18 with freiburg2 and 20 with freiburg3, totalling 48m 12s of recording, 81,413 color frames and 80,683 depth frames.

• Curated by: Computer Vision Group, Technical University of Munich (source release)
• Funded by: [More Information Needed]
• Shared by: Voxel51 (FiftyOne conversion)
• Language(s): Not applicable (sensor data)
• License: CC BY 4.0

Dataset Sources

• Repository: TUM RGB-D SLAM Dataset and Benchmark (source release); Voxel51/TUM-RGBD (this conversion)
• Paper: A Benchmark for the Evaluation of RGB-D SLAM Systems (IROS 2012)
• Demo: [More Information Needed]

Uses

Direct Use

RGB-D SLAM and visual odometry, evaluated against the motion-capture ground truth shipped with each sequence, and work with the registered color and depth images.

Out-of-Scope Use

[More Information Needed]

Dataset Structure

Topology

An ungrouped FiftyOne dataset with media_type="multimodal". One sample is one sequence, and its filepath is a .fo.mcap file. There are 47 samples (9 freiburg1, 18 freiburg2, 20 freiburg3) in six categories, no sample tags, and an empty dataset.info. The dataset has no FiftyOne label fields; every signal lives inside the MCAP file as a channel and is shown by the App's multimodal viewer.

Sample fields

Field	FiftyOne type	Description
id, filepath, tags, metadata, created_at, last_modified_at	built-in	Standard FiftyOne sample fields
sequence	StringField	Sequence name, e.g. fr1/360
camera	StringField	The Kinect that recorded the sequence: freiburg1, freiburg2 or freiburg3
category	StringField	The benchmark's category for the sequence, e.g. Handheld SLAM
recorded	StringField	Recording date
duration	FloatField	Sequence length in seconds
num_rgb_frames	IntField	Color frames
num_depth_frames	IntField	Depth frames
num_accelerometer_samples	IntField	Accelerometer samples; 0 on the freiburg3 sequences
num_ground_truth_poses	IntField	Motion-capture poses of the Kinect
trajectory_length_m	FloatField	Trajectory length in metres, as listed on the benchmark's download page for the sequence
valid_depth_fraction	FloatField	Share of depth pixels holding a measurement
mean_depth_m	FloatField	Mean of the measured depth pixels, in metres

Episode contents

Each MCAP episode contains these channels:

Channel	Schema or content
/rgb	Color images, foxglove.CompressedVideo
/depth	Depth images, foxglove.CompressedImage (16-bit PNG, 5000 per metre)
/rgb-calibration, /depth-calibration	Intrinsics beside each image channel, foxglove.CameraCalibration
/accelerometer.plot	The Kinect's accelerometer, on the freiburg1 and freiburg2 sequences only
/ground-truth	Motion-capture pose of the Kinect, foxglove.PoseInFrame
/ground-truth.plot	Position of the ground-truth pose as a plot
/sequence	Names the sequence and its category

Categories

Category	Sequences	Duration	Color frames
Testing and Debugging	fr1/rpy, fr1/xyz, fr2/rpy, fr2/xyz	4m 43s	8,480
Handheld SLAM	fr1/360, fr1/desk, fr1/desk2, fr1/floor, fr1/room, fr2/360_hemisphere, fr2/360_kidnap, fr2/desk, fr2/large_no_loop, fr2/large_with_loop, fr3/long_office_household	12m 48s	22,864
Robot SLAM	fr2/pioneer_360, fr2/pioneer_slam, fr2/pioneer_slam2, fr2/pioneer_slam3	7m 34s	8,803
Structure vs. Texture	fr3/nostructure_notexture_far, fr3/nostructure_notexture_near_withloop, fr3/nostructure_texture_far, fr3/nostructure_texture_near_withloop, fr3/structure_notexture_far, fr3/structure_notexture_near, fr3/structure_texture_far, fr3/structure_texture_near	4m 19s	7,679
Dynamic Objects	fr2/desk_with_person, fr3/sitting_halfsphere, fr3/sitting_rpy, fr3/sitting_static, fr3/sitting_xyz, fr3/walking_halfsphere, fr3/walking_rpy, fr3/walking_static, fr3/walking_xyz	6m 33s	11,544
3D Object Reconstruction	fr1/plant, fr1/teddy, fr2/coke, fr2/dishes, fr2/flowerbouquet, fr2/flowerbouquet_brownbackground, fr2/metallic_sphere, fr2/metallic_sphere2, fr3/cabinet, fr3/large_cabinet, fr3/teddy	12m 16s	22,043

Parsing decisions

• Why MCAP: the dataset is one multimodal sample per sequence, so the color and depth images, the accelerometer and the ground-truth poses play back on one shared clock instead of being split into per-frame samples.
• Depth: the depth images are carried as the benchmark ships them, 16-bit PNG in which 5000 is one metre and 0 is no measurement, pre-registered to the color image so that their pixels correspond one to one.
• Video encoding: the color images are re-encoded to Annex-B H.264 without B-frames, one access unit per frame, on each image's own timestamp.
• Calibration: both calibration channels carry the benchmark's published intrinsics for the Kinect that recorded the sequence, which the registered depth shares. The freiburg3 images are already undistorted, so their distortion is zero.
• Ground truth: the ground truth is the benchmark's groundtruth.txt, the pose of the Kinect in the motion-capture frame, carried as /ground-truth in a frame named world.
• Accelerometer: the accelerometer is carried as recorded. The benchmark's freiburg3 archives hold no accelerometer readings, so those episodes have none.
• Left out: the validation sequences, whose ground truth the benchmark withholds, and the calibration recordings are not included.

Dataset Creation

Curation Rationale

[More Information Needed]

Source Data

Data Collection and Processing

A Microsoft Kinect records color and depth at 640x480 and 30 Hz while an eight-camera motion-capture system tracks it at 100 Hz. Three Kinects, named freiburg1, freiburg2 and freiburg3 by the benchmark, record the sequences, handheld, on a Pioneer robot, and over scenes built to test structure against texture, moving people and object reconstruction. The FiftyOne conversion reads the benchmark's image archives and writes one MCAP episode per sequence; see Parsing decisions above for the changes made.

Who are the source data producers?

The Computer Vision Group at the Technical University of Munich.

Annotations

Annotation process

The dataset has no human annotations. The ground truth in each sequence is the pose of the Kinect recorded by the eight-camera motion-capture system at 100 Hz, as listed in the benchmark's groundtruth.txt.

Who are the annotators?

Not applicable.

Personal and Sensitive Information

The Dynamic Objects category contains sequences with people moving through the scene. [More Information Needed]

Citation

The benchmark asks that work using the data cite:

BibTeX:

@InProceedings{sturm12iros,
  author = {J. Sturm and N. Engelhard and F. Endres and W. Burgard and D. Cremers},
  title = "A Benchmark for the Evaluation of RGB-D SLAM Systems",
  booktitle = "Proc. of the International Conference on Intelligent Robot Systems (IROS)",
  year = "2012",
  month = "Oct."
}

APA:

Sturm, J., Engelhard, N., Endres, F., Burgard, W., & Cremers, D. (2012, October). A benchmark for the evaluation of RGB-D SLAM systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS).

More Information

The TUM RGB-D benchmark's data is distributed under the Creative Commons Attribution 4.0 International license (CC-BY-4.0), and this conversion is distributed under the same license.

Changes from the source: conversion from the benchmark's image archives to the FiftyOne MCAP flavor, H.264 encoding of the color images, the benchmark's published intrinsics carried as camera calibration, and the validation and calibration recordings left out.

Dataset Card Authors

[More Information Needed]

Dataset Card Contact

[More Information Needed]