Giters

adrianbojko / consinv-dataset

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

ConsInv Dataset

This repository contains the dataset corresponding to the BMVC 2022 paper "Self-Improving SLAM in Dynamic Environments: Learning When to Mask" with train/val/test splits. The paper itself, supplementary materials, poster and video are here: https://bmvc2022.mpi-inf.mpg.de/654/ . ConsInv dataset includes 150 lossless sequences and over 85,000 pairs of stereo images!

ConsInv is a stereo RGB + IMU dataset designed for Dynamic SLAM testing and contains two subsets:

  • ConsInv-Indoors contains sequences in an office setting where small objects are moved.
  • ConsInv-Outdoors contains sequences in an urban environment, where cars and/or people move.

The novelty of ConsInv dataset is 1) the controlled degree of difficulty, from easy to very hard, and 2) the fact that the difficulty of the sequences comes only from object motion: relative motion between camera and object, motion ambiguity, challenging points of view when objects move. The difficulty does not come from motion speed, lack of features, lens flare, etc. - typically seen in other SLAM datasets.

ConsInv Dataset

ConsInv-Outdoors (top) and ConsInv-Indoors (bottom)

Citation

    @inproceedings{Bojko_2022_BMVC,
    author    = {Adrian Bojko and Romain Dupont and Mohamed Tamaazousti and Herve Le Borgne},
    title     = {Self-Improving SLAM in Dynamic Environments: Learning When to Mask},
    booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
    publisher = {{BMVA} Press},
    year      = {2022},
    url       = {https://bmvc2022.mpi-inf.mpg.de/0654.pdf}
    }

Download

Calibration files here:

ConsInv-Indoors is under CC BY-NC-SA 4.0 license and can be downloaded here:

ConsInv-Outdoors is under a different license, similar to CC BY-NC-SA 4.0. It requires personal identifying information (e.g., human faces, car registration plates) to be removed when content from the database is shared, e.g., in papers or online. To obtain a free access to this subset, please contact us. You can get two anonymized sequences here:

TUM RGB-D semantic masks (16.3 MB) used in our paper and computed with the model from our other paper Learning to Segment Dynamic Objects using SLAM Outliers are available here:

Contact

Please contact us at consinv-dataset{@}bojko.ai . Alternatively, for issues other than access to ConsInv-Outdoors, you can open a Git issue.

Notes

General

  • Onedrive may ask for login with a Microsoft account to download large files. We have no control over this as it is an imposed abuse protection from Microsoft. You will be able to download files immediately after login; you do not need to ask for permissions.
  • The dataset is huge due to having 1280x720 stereo lossless PNG files.
  • The anonymized ConsInv-Outdoors samples can be used to share images/videos without having to anonymize them.

Datasets

  • The general naming pattern of sequences is (tr|v|te)_[static_env_|static_cam_]_(easy|hard|veryhard)_*. tr/v/te respectively indicate training/validation/test sequences. static_cam indicates that the camera does not move and static_env that no objects move. easy/hard/veryhard gives a qualitative estimate of the difficulty of sequences in terms of dynamic motion of both camera and objects.
  • We provide monocular and stereo semantic masks for ConsInv-Outdoors. Stereo semantic masks are computed on rectified stereo images.
  • delay.txt (global delay) / delay_maskrcnn.txt (per-class delays, only some sequences) give the frames numbers from which you should start applying semantic masks to get a good baseline without a full Temporal Masking pipeline.
  • ConsInv-Indoors is designed for monocular use and ConsInv-Outdoors for stereo use. Still, we provide stereo images for both as well as IMU data. Please note that IMU data has not been used nor checked.
  • Some sequences may seem to have a "washed" color. This is normal and due to the camera type (Mynt EYE D1000-120).
  • The ground truth has been checked to be consistent with real-world measurements. Still, it was not computed with a motion capture system and may not be perfect. Ground truth files are not provided for static_cam sequences as the camera is static, i.e., the ground truth is a constant zero.

About