Voxel51/TUM-RGBD

TUM RGB-D SLAM benchmark → FiftyOne (Native Multimodal MCAP)

The TUM RGB-D SLAM Dataset and Benchmark from the Computer Vision Group at the Technical University of Munich, converted to native multimodal MCAP episodes.

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.

48m 12s of recording, 81,413 color frames and 80,683 depth frames.

Installation

pip install fiftyone

Usage

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

dataset = fouh.load_from_hub(
    "Voxel51/TUM-RGBD",
    name="TUM-RGBD",
    persistent=True,
)
fo.launch_app(dataset)

The sequences with people moving through the scene:

view = dataset.match(F("category") == "Dynamic Objects")
fo.launch_app(dataset, view=view)

with from fiftyone import ViewField as F.

What you get

Each episode contains:

• /rgb, the color images, as foxglove.CompressedVideo
• /depth, the depth images, as foxglove.CompressedImage (16-bit PNG, 5000 per metre)
• a -calibration topic beside each, as foxglove.CameraCalibration
• /accelerometer.plot, the Kinect's accelerometer, on the freiburg1 and freiburg2 sequences
• /ground-truth, the motion-capture pose of the Kinect, as foxglove.PoseInFrame, with its position on /ground-truth.plot
• /sequence, naming the sequence and its category

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

Episodes carry the fields sequence, camera, category, recorded, duration, num_rgb_frames, num_depth_frames, num_accelerometer_samples, num_ground_truth_poses, trajectory_length_m, valid_depth_fraction and mean_depth_m. trajectory_length_m is the length the benchmark's download page lists for the sequence, valid_depth_fraction the share of depth pixels holding a measurement, and mean_depth_m their mean.

Notes on the conversion

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. 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. Both calibration topics 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.

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. The accelerometer is carried as recorded; the benchmark's freiburg3 archives hold no accelerometer readings, so those episodes have none. The validation sequences, whose ground truth the benchmark withholds, and the calibration recordings are not included.

License & attribution

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. The benchmark asks that work using the data cite:

@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."
}

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.