Access TOPO-Bench

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

TOPO-Bench repackages sequences from OpenLORIS-Scene, RELLIS-3D, ROVER and Rawseeds, and RGB-D images rendered from HM3D scenes (Matterport). Before downloading, you must have accepted the Matterport terms of use for HM3D (https://matterport.com/habitat-matterport-3d-research-dataset). You may use this dataset only for non-commercial research, under the license of each source dataset listed in this card.

Log in or Sign Up to review the conditions and access this dataset content.

TOPO-Bench: topological localization under perceptual aliasing

TOPO-Bench tests whether a robot that has mapped a place can tell where it is after being moved without knowing it, when other places in the map look alike. A system builds a map from a mapping traversal. It is then dropped into a test traversal without odometry for the first frame, follows it, and at the last test frame must either name the map place it is at or abstain because that place is not in the map. Each case belongs to one of three classes:

class true place in the map look-alike place in the map correct answer
A+P (high_distraction) yes yes, about as similar as the true place the true place
P.O. (low_distraction) yes no, or clearly less similar the true place
A.O. (no_target) no (removed from the map) yes abstain

It is the benchmark of TOPO-Bench: An Open-Source Topological Mapping Evaluation Framework with Quantifiable Perceptual Aliasing (Wang, Liu, Chen, Da, Hu, Xue, Kästner, Soh; National University of Singapore; project page). The code (evaluator, baselines, CROSS adapter, download and check scripts) is at https://github.com/jiaming-ai/topo-bench-release.

Contents (v1.1)

dataset scenes / sequences A+P P.O. A.O. sensor data acceptance size (full / lite)
OpenLORIS-Scene 5 / 21 30 70 52 RealSense D435i RGB-D 848×480, wheel odometry 2 m 59 GB / 4 GB
Habitat (HM3D renders) 15 / 30 0 65 2 RGB-D 1280×720 (fx 605 px, measured from the images), noisy synthetic odometry 2 m 14 GB / 1 GB
RELLIS-3D 1 / 5 0 0 117 RGB 1920×1200 with depth (≤ 20 m), odometry 2 m 21 GB / 1 GB
ROVER (campus_large) 1 / 5 5 30 8 RealSense D435i RGB-D 640×480 (radtan distortion), odometry; five recordings over a year 5 m 86 GB / 8 GB
total 35 165 179 181 GB / 14 GB

Every frame comes with a ground-truth pose, the odometry increment since the previous frame, and its BoQ (ResNet-50) place-recognition embedding, so embedding-based methods run without images (the lite download).

Scoring

The evaluator streams the case's map (for A.O. cases only the map fragments that remain after removing the true place), then the test traversal from 200 frames before its start to its end; the first test frame has no odometry. Only the answer at the last test frame counts. For A+P and P.O. cases it is correct when it names one of the case's acceptable_mapping_frames: the map frames within the case's acceptance.radius_m (2 m; ROVER 5 m) and acceptance.angle_deg (120°) of the last test frame by ground truth. For A.O. cases it is correct when the system abstains. The per-class accuracies are combined into BLA, the geometric mean of the three accuracies, each smoothed as (k + 0.5) / (n + 1). Systems with a confidence threshold are compared over all thresholds (L_AO@90 and BLA_max in the paper); scripts/sweep_thresholds.py in the code computes both.

How the cases were built

These steps follow the scripts that generated the cases. They differ from the paper's text in one place: the class label compares the distractor with the true place (step 4) instead of applying an absolute similarity threshold.

  1. Keyframes: a new keyframe whenever the BoQ cosine similarity to the previous keyframe drops to 0.8 or below.
  2. Overlap (the true place): a run of test keyframes whose best-matching map keyframe has similarity at least 0.35 (at least 0.50: a strong match) and lies within 1.5 m and 90° of it by ground truth.
  3. Distractor: a run of more than three consecutive map keyframes that match consecutive test keyframes with similarity at least 0.35, all more than 10 m from the overlap.
  4. Class: without a distractor, P.O. With one, A+P when the mean similarity of the distractor's last two keyframes exceeds 0.9 times that of the overlap's last two, else P.O. In addition, a third (fixed seed) of the cases with a strong match and a distractor weaker at each of the last two keyframes are copied into A.O. cases, whose map loses the overlap and every frame within 3 m of it.
  5. Acceptable frames: as in Scoring, recomputed for v1.1 from the ground truth.

Layout

cases/<dataset>/<scene>.json                  the cases of a scene
sequences/<dataset>/<scene>/<sequence>.meta.tar  everything but the images (unpacks into <scene>/<sequence>/):
    rgb.txt or color.txt     the images in time order: <timestamp> <relative path>
    aligned_depth.txt        the depth map of each image: <image timestamp> <relative path>
    groundtruth.txt          <timestamp> <x> <y> <heading> (OpenLORIS also: ground_truth_poses.pkl, 4x4 poses)
    aligned_odometry.pkl     one 4x4 increment per image, inv(T_prev) @ T_curr in the OpenCV camera frame
                             (x right, y down, z forward); rotations orthonormal
    camera.json              K, width, height, depth_scale (depth units per metre), depth_max (depth at or above it
                             is invalid), distortion (OpenCV [k1, k2, p1, p2], ROVER only), odometry_frame
    boq.npz                  BoQ (ResNet-50) embedding of every image
sequences/<dataset>/<scene>/<sequence>.images.tar[.NNN]  images and depth maps (parts of at most ~20 GB)
manifest.json                 version, files per dataset, case counts, file sizes
SHA256SUMS                    checksum of every file above

A case names frames by path relative to its dataset folder (<scene>/<sequence>/<image folder>/<file>): start_mapping / end_mapping (the map; A.O. cases list their fragments in mapping_sections), start_testing / end_testing (the test span), start_overlap / end_overlap (where the true place is in the map), start_ambiguous / end_ambiguous (the distractor), acceptable_mapping_frames, acceptance, class and a stable index. *.pkl files are Python pickles: load them only from this repository.

Download

Click Agree and access on this page and log in with hf auth login. Then, from a clone of the code:

python scripts/download_data.py                        # all datasets (~181 GB)
python scripts/download_data.py --lite                 # no images (~14 GB): enough for GM / SM / PBU
python scripts/download_data.py --datasets rellis      # some datasets only
python scripts/check_data.py                           # every case resolves; counts match the release

The script downloads one tagged release (v1.1 by default), checks every file against SHA256SUMS and unpacks it into data/<dataset>/<scene>/. Without the code: hf download Jamie-nus/TOPO-Bench --repo-type dataset --revision v1.1 --local-dir TOPO-Bench, unpack every archive of sequences/<dataset>/ into data/<dataset>/, and copy cases/<dataset>/<scene>.json to data/<dataset>/<scene>/cases.json.

Versions

  • v1.1 (default): the corrected benchmark below.
  • v1.0: the case files as published with the paper (--revision v1.0), with the same sequence data plus Rawseeds, for reproducing the paper's numbers. Its errors are listed below; prefer v1.1 for new results.

Changes from v1.0 to v1.1:

  • ROVER: the v1.0 case file held one of three parts of a split, 14 of the 43 cases; v1.1 has all 43.
  • Rawseeds removed. Its test sequences are cuts of whole recordings, each moved into a scene's map frame by a hand-picked transform. Cuts of a recording into its own map are 2–4 m from where they belong, 15 cases score a test frame that is also in the map, and we could not find one registration that holds for all cuts, so its 2 m acceptable lists cannot be trusted.
  • Removed 4 duplicate cases each in OpenLORIS and Habitat, and one OpenLORIS A.O. case whose map was empty.
  • Acceptable lists recomputed from the ground truth with one rule (step 5). The v1.0 Habitat lists ignored the heading (a bug in their generator) and 19 of 65 change; the OpenLORIS lists are unchanged; 21 of the 35 ROVER lists change slightly.
  • Sensor metadata for systems that use images and odometry: camera.json per sequence, odometry as OpenCV-frame increments in every dataset (the sources use three conventions), ground truth as t x y heading (the RELLIS-3D and ROVER sources carry extra columns that the v1.0 loader misread), depth paths relative.
  • Class labels are unchanged.

Not included: Habitat scene 00305 (its case file was never finished and has no acceptable lists; v1.0 skipped it too) and Oxford RobotCar, whose authors withdrew the dataset on 2026-10-22.

License

Each sequence stays under the license of its source; the benchmark definition (cases, acceptable lists, normalised metadata and embeddings) is released for non-commercial research use.

source terms
OpenLORIS-Scene CC BY-ND 4.0; this processed subset is redistributed with the authors' permission
Habitat renders of HM3D Matterport terms of use for HM3D: non-commercial research only, no redistribution
RELLIS-3D CC BY-NC-SA 3.0
ROVER the ROVER dataset terms (listed as MIT)
Rawseeds (v1.0 only) Rawseeds project terms; redistributed with permission

Citation

@inproceedings{wang2026topobench,
  title     = {{TOPO-Bench}: An Open-Source Topological Mapping Evaluation Framework with Quantifiable Perceptual Aliasing},
  author    = {Wang, Jiaming and Liu, Diwen and Chen, Jizhuo and Da, Jiaxuan and Hu, Jiamo and Xue, Zhiwei and K{\"a}stner, Linh and Soh, Harold},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026},
  url       = {https://clear-nus.github.io/topo-bench/}
}

Please also cite the source datasets and BoQ:

@inproceedings{shi2019openlorisscene,
  title     = {Are We Ready for Service Robots? The {OpenLORIS-Scene} Datasets for Lifelong {SLAM}},
  author    = {Shi, Xuesong and Li, Dongjiang and Zhao, Pengpeng and Tian, Qinbin and Tian, Yuxin and Long, Qiwei and Zhu, Chunhao and Song, Jingwei and Qiao, Fei and Song, Le and Guo, Yangquan and Wang, Zhigang and Zhang, Yimin and Qin, Baoxing and Yang, Wei and Wang, Fangshi and Chan, Rosa H. M. and She, Qi},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  pages     = {3139--3145},
  year      = {2020}
}
@inproceedings{habitat19iccv,
  title     = {Habitat: A Platform for Embodied {AI} Research},
  author    = {Savva, Manolis and Kadian, Abhishek and Maksymets, Oleksandr and Zhao, Yili and Wijmans, Erik and Jain, Bhavana and Straub, Julian and Liu, Jia and Koltun, Vladlen and Malik, Jitendra and Parikh, Devi and Batra, Dhruv},
  booktitle = {IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2019}
}
@article{jiang2020rellis3d,
  title   = {{RELLIS-3D} Dataset: Data, Benchmarks and Analysis},
  author  = {Jiang, Peng and Osteen, Philip and Wigness, Maggie and Saripalli, Srikanth},
  journal = {arXiv preprint arXiv:2011.12954},
  year    = {2020}
}
@article{schmidt2025rover,
  title   = {{ROVER}: A Multiseason Dataset for Visual {SLAM}},
  author  = {Schmidt, Fabian and Daubermann, Julian and Mitschke, Marcel and Blessing, Constantin and Meyer, Stephan and Enzweiler, Markus and Valada, Abhinav},
  journal = {IEEE Transactions on Robotics},
  volume  = {41},
  pages   = {4005--4022},
  year    = {2025}
}
@incollection{fontana2013rawseeds,
  title     = {Rawseeds: Building a Benchmarking Toolkit for Autonomous Robotics},
  author    = {Fontana, Giulio and Matteucci, Matteo and Sorrenti, Domenico G.},
  booktitle = {Methods and Experimental Techniques in Computer Engineering},
  pages     = {55--68},
  publisher = {Springer},
  year      = {2013}
}
@inproceedings{ali2024boq,
  title     = {{BoQ}: A Place is Worth a Bag of Learnable Queries},
  author    = {Ali-Bey, Amar and Chaib-draa, Brahim and Gigu{\`e}re, Philippe},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {17794--17803},
  year      = {2024}
}
Downloads last month
1,155

Paper for Jamie-nus/TOPO-Bench