Emulated-Inc/scenepose

Scene Camera Pose from Unordered Image Collections

Multi-view image collections of thirty distinct scenes, landmark photo collections and systematic captures of sites and objects. Training scenes carry a ground-truth camera pose per image and per-scene registration thresholds; held-out scenes carry images only.

Contents

The data tree lives under data/, exactly as the benchmark environment presents it at /app/scenepose. Its own description is at data/README.md.

• 1482 files, 1031106944 bytes (1.0 GiB)
• structure hash (sha256 over sorted path\tsize lines): bfe5460bdc29882ad1e7518417db2931b634f234edf3f8cb1cbee299f14ec3cc
• data.manifest.tsv at the repo root lists every file as sha256 size path

Use

from huggingface_hub import snapshot_download

snapshot_download("Emulated-Inc/scenepose", repo_type="dataset", revision="REVISION",
                  local_dir="./scenepose", allow_patterns="data/*")

Pin revision to a commit sha rather than a branch if you need reproducibility.

Provenance and licence

Derived from the training data of the Kaggle Image Matching Challenge 2024 (https://www.kaggle.com/competitions/image-matching-challenge-2024) and Image Matching Challenge 2025 (https://www.kaggle.com/competitions/image-matching-challenge-2025), both released under the Apache License 2.0 by the Image Matching Challenge organisers (Czech Technical University in Prague and collaborators), including PhotoTourism landmark collections derived from YFCC100M. Scenes and images are renamed to opaque codes, JPEG re-encoded with metadata stripped, and subsampled per scene. The held-out split used by the benchmark this dataset serves is carved from that training data; its poses are not included here.