Articulate3D Challenge

🚀 Getting Started

1. Get the Data

You need two things: Articulate3D annotations and the ScanNet++ meshes they're defined on top of.

Annotations (train + val splits) — huggingface.co/datasets/INSAIT-Institute/Articulate3D
Meshes — request access and download from scannetpp.mlsg.cit.tum.de/scannetpp. You'll receive a download_scannetpp.py toolkit + access token from them.

Which scenes should I download? (Articulate3D Splits)

⚠️ Articulate3D's train/val/test split is not the same as ScanNet++'s own. Don't download the entire ScanNet++ dataset — it's huge and unnecessary. Instead, only download the scenes with ids provided in the lists below. These are the scenes that Articulate3D uses for training, validation, and testing.
Articulate3D remixes scenes from ScanNet++ v1's train and val sets into its own train/val/test split. Always use these scene ID lists — not ScanNet++'s own split files — or you'll train/evaluate on the wrong scenes.

📂 View scene ID lists (train / val / test)

TRAIN_SCENES = [
    "f248c2bcdc", "94ee15e8ba", "d918af9c5f", "6ee2fc1070", "cf1ffd871d",
    "9f21bdec45", "e0de253456", "7079b59642", "394a542a19", "32280ecbca",
    "c50d2d1d42", "f25f5e6f63", "acd69a1746", "54b6127146", "320c3af000",
    "6b40d1a939", "50809ea0d8", "dfac5b38df", "e01b287af5", "7cd2ac43b4",
    "7f4d173c9c", "4318f8bb3c", "d415cc449b", "5942004064", "8e6ff28354",
    "825d228aec", "13285009a4", "6115eddb86", "3f15a9266d", "07ff1c45bb",
    "7977624358", "c856c41c99", "d7abfc4b17", "d6702c681d", "280b83fcf3",
    "bb87c292ad", "37ea1c52f0", "076c822ecc", "56a0ec536c", "7e09430da7",
    "c9abde4c4b", "ab6983ae6c", "13c3e046d7", "6cc2231b9c", "b1d75ecd55",
    "31a2c91c43", "cc5237fd77", "85251de7d1", "a1d9da703c", "ab046f8faf",
    "40aec5fffa", "9f79564dbf", "acd95847c5", "f07340dfea", "cbd4b3055e",
    "893fb90e89", "ab11145646", "8b2c0938d6", "116456116b", "75d29d69b8",
    "dc263dfbf0", "4c5c60fa76", "7eac902fd5", "c545851c4f", "e0abd740ba",
    "d6d9ddb03f", "961911d451", "303745abc7", "c8f2218ee2", "5fb5d2dbf2",
    "f3685d06a9", "260db9cf5a", "7b6477cb95", "ed2216380b", "3e8bba0176",
    "f8f12e4e6b", "e9ac2fc517", "f9f95681fd", "260fa55d50", "410c470782",
    "8a35ef3cfe", "28a9ee4557", "ebc200e928", "d2f44bf242", "c173f62b15",
    "709ab5bffe", "95d525fbfd", "e8e81396b6", "98fe276aa8", "aaa11940d3",
    "302a7f6b67", "9471b8d485", "08bbbdcc3d", "e1b1d9de55", "f2dc06b1d2",
    "07f5b601ee", "3db0a1c8f3", "2b1dc6d6a5", "09bced689e", "a4e227f506",
    "4ba22fa7e4", "3c95c89d61", "ef25276c25", "210f741378", "fe1733741f",
    "8be0cd3817", "80ffca8a48", "0a7cc12c0e", "281bc17764", "6d89a7320d",
    "f34d532901", "25927bb04c", "d755b3d9d8", "1c4b893630", "c5f701a8c7",
    "1831b3823a", "9f139a318d", "286b55a2bf", "5d152fab1b", "0cf2e9402d",
    "d6cbe4b28b", "824d9cfa6e", "99fa5c25e1", "8133208cb6", "39f36da05b",
    "1d003b07bd", "0a5c013435", "1a8e0d78c0", "fd361ab85f", "ef69d58016",
    "e050c15a8d", "689fec23d7", "0a184cf634", "104acbf7d2", "9460c8889d",
    "16c9bd2e1e", "290ef3f2c9", "281ba69af1", "3864514494", "2970e95b65",
    "8e00ac7f59", "079a326597", "e8ea9b4da8", "5f99900f09", "eb4bc76767",
    "0b031f3119", "fb5a96b1a2", "e3ecd49e2b", "30966f4c6e", "45b0dac5e3",
    "1ada7a0617", "89214f3ca0", "0e75f3c4d9", "ef18cf0708", "712dc47104",
    "0d2ee665be", "09c1414f1b", "355e5e32db", "49a82360aa", "39e6ee46df",
    "578511c8a9", "3928249b53", "e7af285f7d", "ccfd3ed9c7", "2a496183e1",
    "646af5e14b", "87f6d7d564", "8b5caf3398", "6f1848d1e3", "fb05e13ad1",
    "88cf747085", "3e928dc2f6", "b0a08200c9", "47b37eb6f9", "daffc70503",
    "108ec0b806", "7bc286c1b6", "f6659a3107", "480ddaadc0", "3f1e1610de",
    "484ad681df", "c4c04e6d6c", "7831862f02", "6855e1ac32", "69e5939669",
    "7e7cd69a59", "c0f5742640", "faec2f0468", "5a269ba6fe", "a05ee63164",
    "59e3f1ea37", "251443268c", "036bce3393", "4ea827f5a1", "21d970d8de",
]

VAL_SCENES = [
    "8f82c394d6", "5656608266", "1ae9e5d2a6", "a980334473", "27dd4da69e",
    "88627b561e", "8d563fc2cc", "e91722b5a3", "4422722c49", "40b56bf310",
    "38d58a7a31", "e398684d27", "e9e16b6043", "b20a261fdf", "0a76e06478",
    "5eb31827b7", "bcd2436daf", "f8062cb7ce", "5654092cc2", "30f4a2b44d",
    "25f3b7a318", "1204e08f17", "9859de300f", "6464461276", "324d07a5b3",
    "2e74812d00", "41b00feddb", "f3d64c30f8", "1366d5ae89", "8a20d62ac0",
    "a29cccc784", "55b2bf8036", "1841a0b525", "e898c76c1f", "61adeff7d5",
    "5ee7c22ba0", "8890d0a267", "9071e139d9", "4a1a3a7dc5", "5748ce6f01",
    "419cbe7c11", "f5401524e5",
]

TEST_SCENES = [
    "98b4ec142f", "c49a8c6cff", "c0c863b72d", "ac48a9b736", "b5918e4637",
    "b08a908f0f", "b26e64c4b0", "785e7504b9", "bd7375297e", "c413b34238",
    "1b9692f0c7", "bf6e439e38", "9b74afd2d2", "67d702f2e8", "a24f64f7fb",
    "ad2d07fd11", "a08d9a2476", "a08dda47a8", "bc03d88fc3", "1f7cbbdde1",
    "1a130d092a", "b97261909e", "bc400d86e1", "bd9305480d", "bde1e479ad",
    "52599ae063", "c06a983e63", "b09431c547", "a003a6585e", "be0ed6b33c",
    "66c98f4a9b", "c5439f4607", "e7ac609391", "1b75758486", "ada5304e41",
    "b73f5cdc41", "c24f94007b", "bc2fce1d81", "a5114ca13d", "c47168fab2",
    "a8bf42d646", "bfd3fd54d2",
]

195 train / 42 val / 42 test scenes. Test-split annotations are not published (held out for scoring) — you only get meshes for those 42. Training on train + val combined is fine; final leaderboard standing is on test only.

Downloading the Meshes

You don't need all per-scene assets for the relevant scenes, just the aligned mesh (mesh_aligned_0.05.ply) for each scene. Use the script below with the ScanNet++ toolkit (scene_release.py must be in the same folder — it ships with the toolkit). It downloads just the mesh (mesh_aligned_0.05.ply) for each scene, by scene ID directly. You can also use the scannetpp toolkit's download_scannetpp.py and pass them the scene IDs, but this script is simpler and downloads only the meshes you need.

Our script for downloading only the meshes for the relevant scenes (Python). You need to provide your ScanNet++ access token in my_config.yml (see below).

# save as download_meshes.py, next to the ScanNet++ toolkit's scene_release.py
import sys, zipfile
from pathlib import Path
sys.path.insert(0, str(Path(__file__).parent))
from scene_release import ScannetppScene_Release
from download_scannetpp import check_download_file, load_yaml_munch

CFG_PATH = "my_config.yml"   # token + data_root, see below

TRAIN_SCENES = [
    "f248c2bcdc", "94ee15e8ba", "d918af9c5f", "6ee2fc1070", "cf1ffd871d",
    "9f21bdec45", "e0de253456", "7079b59642", "394a542a19", "32280ecbca",
    "c50d2d1d42", "f25f5e6f63", "acd69a1746", "54b6127146", "320c3af000",
    "6b40d1a939", "50809ea0d8", "dfac5b38df", "e01b287af5", "7cd2ac43b4",
    "7f4d173c9c", "4318f8bb3c", "d415cc449b", "5942004064", "8e6ff28354",
    "825d228aec", "13285009a4", "6115eddb86", "3f15a9266d", "07ff1c45bb",
    "7977624358", "c856c41c99", "d7abfc4b17", "d6702c681d", "280b83fcf3",
    "bb87c292ad", "37ea1c52f0", "076c822ecc", "56a0ec536c", "7e09430da7",
    "c9abde4c4b", "ab6983ae6c", "13c3e046d7", "6cc2231b9c", "b1d75ecd55",
    "31a2c91c43", "cc5237fd77", "85251de7d1", "a1d9da703c", "ab046f8faf",
    "40aec5fffa", "9f79564dbf", "acd95847c5", "f07340dfea", "cbd4b3055e",
    "893fb90e89", "ab11145646", "8b2c0938d6", "116456116b", "75d29d69b8",
    "dc263dfbf0", "4c5c60fa76", "7eac902fd5", "c545851c4f", "e0abd740ba",
    "d6d9ddb03f", "961911d451", "303745abc7", "c8f2218ee2", "5fb5d2dbf2",
    "f3685d06a9", "260db9cf5a", "7b6477cb95", "ed2216380b", "3e8bba0176",
    "f8f12e4e6b", "e9ac2fc517", "f9f95681fd", "260fa55d50", "410c470782",
    "8a35ef3cfe", "28a9ee4557", "ebc200e928", "d2f44bf242", "c173f62b15",
    "709ab5bffe", "95d525fbfd", "e8e81396b6", "98fe276aa8", "aaa11940d3",
    "302a7f6b67", "9471b8d485", "08bbbdcc3d", "e1b1d9de55", "f2dc06b1d2",
    "07f5b601ee", "3db0a1c8f3", "2b1dc6d6a5", "09bced689e", "a4e227f506",
    "4ba22fa7e4", "3c95c89d61", "ef25276c25", "210f741378", "fe1733741f",
    "8be0cd3817", "80ffca8a48", "0a7cc12c0e", "281bc17764", "6d89a7320d",
    "f34d532901", "25927bb04c", "d755b3d9d8", "1c4b893630", "c5f701a8c7",
    "1831b3823a", "9f139a318d", "286b55a2bf", "5d152fab1b", "0cf2e9402d",
    "d6cbe4b28b", "824d9cfa6e", "99fa5c25e1", "8133208cb6", "39f36da05b",
    "1d003b07bd", "0a5c013435", "1a8e0d78c0", "fd361ab85f", "ef69d58016",
    "e050c15a8d", "689fec23d7", "0a184cf634", "104acbf7d2", "9460c8889d",
    "16c9bd2e1e", "290ef3f2c9", "281ba69af1", "3864514494", "2970e95b65",
    "8e00ac7f59", "079a326597", "e8ea9b4da8", "5f99900f09", "eb4bc76767",
    "0b031f3119", "fb5a96b1a2", "e3ecd49e2b", "30966f4c6e", "45b0dac5e3",
    "1ada7a0617", "89214f3ca0", "0e75f3c4d9", "ef18cf0708", "712dc47104",
    "0d2ee665be", "09c1414f1b", "355e5e32db", "49a82360aa", "39e6ee46df",
    "578511c8a9", "3928249b53", "e7af285f7d", "ccfd3ed9c7", "2a496183e1",
    "646af5e14b", "87f6d7d564", "8b5caf3398", "6f1848d1e3", "fb05e13ad1",
    "88cf747085", "3e928dc2f6", "b0a08200c9", "47b37eb6f9", "daffc70503",
    "108ec0b806", "7bc286c1b6", "f6659a3107", "480ddaadc0", "3f1e1610de",
    "484ad681df", "c4c04e6d6c", "7831862f02", "6855e1ac32", "69e5939669",
    "7e7cd69a59", "c0f5742640", "faec2f0468", "5a269ba6fe", "a05ee63164",
    "59e3f1ea37", "251443268c", "036bce3393", "4ea827f5a1", "21d970d8de",
]

VAL_SCENES = [
    "8f82c394d6", "5656608266", "1ae9e5d2a6", "a980334473", "27dd4da69e",
    "88627b561e", "8d563fc2cc", "e91722b5a3", "4422722c49", "40b56bf310",
    "38d58a7a31", "e398684d27", "e9e16b6043", "b20a261fdf", "0a76e06478",
    "5eb31827b7", "bcd2436daf", "f8062cb7ce", "5654092cc2", "30f4a2b44d",
    "25f3b7a318", "1204e08f17", "9859de300f", "6464461276", "324d07a5b3",
    "2e74812d00", "41b00feddb", "f3d64c30f8", "1366d5ae89", "8a20d62ac0",
    "a29cccc784", "55b2bf8036", "1841a0b525", "e898c76c1f", "61adeff7d5",
    "5ee7c22ba0", "8890d0a267", "9071e139d9", "4a1a3a7dc5", "5748ce6f01",
    "419cbe7c11", "f5401524e5",
]

TEST_SCENES = [
    "98b4ec142f", "c49a8c6cff", "c0c863b72d", "ac48a9b736", "b5918e4637",
    "b08a908f0f", "b26e64c4b0", "785e7504b9", "bd7375297e", "c413b34238",
    "1b9692f0c7", "bf6e439e38", "9b74afd2d2", "67d702f2e8", "a24f64f7fb",
    "ad2d07fd11", "a08d9a2476", "a08dda47a8", "bc03d88fc3", "1f7cbbdde1",
    "1a130d092a", "b97261909e", "bc400d86e1", "bd9305480d", "bde1e479ad",
    "52599ae063", "c06a983e63", "b09431c547", "a003a6585e", "be0ed6b33c",
    "66c98f4a9b", "c5439f4607", "e7ac609391", "1b75758486", "ada5304e41",
    "b73f5cdc41", "c24f94007b", "bc2fce1d81", "a5114ca13d", "c47168fab2",
    "a8bf42d646", "bfd3fd54d2",
]
SCENES = TRAIN_SCENES + VAL_SCENES + TEST_SCENES  # or just the split(s) you need

def main():
    cfg = load_yaml_munch(CFG_PATH)
    data_root = Path(cfg.data_root)
    for i, scene_id in enumerate(SCENES, 1):
        print(f"[{i}/{len(SCENES)}] {scene_id}")
        src = ScannetppScene_Release(scene_id, data_root="data")
        tgt = ScannetppScene_Release(scene_id, data_root=data_root / "data")
        if tgt.scan_mesh_path.is_file():
            continue
        zip_path = tgt.scan_mesh_path.with_suffix(".zip")
        if check_download_file(cfg, cfg.root_url, src.scan_mesh_path.with_suffix(".zip"), zip_path, cfg.dry_run):
            with zipfile.ZipFile(zip_path) as zf:
                zf.extractall(zip_path.parent)
            zip_path.unlink()

if __name__ == "__main__":
    main()

my_config.yml (minimal — just needs your token and where to save):

token: <your ScanNet++ token>
data_root: ./scannetpp_data
root_url: https://scannetpp.mlsg.cit.tum.de/scannetpp/download/v2?version=v1&token=TOKEN&file=FILEPATH
dry_run: false
verbose: true

2. Easiest Way to Start: the Baseline

We provide a ready baseline with data loaders and training/inference scripts — start here instead of building your own pipeline from scratch:

Code: github.com/insait-institute/USDNet
Preprocessed data (skip raw mesh + annotation preprocessing entirely): Drive folder

Drop the preprocessed data into USDNet/data/processed/ per the repo's README and you're ready to train/run inference — this is the point cloud format the challenge server itself scores against, byte-for-byte, so using it also sidesteps any risk of your own preprocessing producing a differently-ordered point cloud.

3. Submission Format

Predictions are a single .pkl file: a dict keyed by scene ID, one entry per scene.

Movable phases (mov_dev, mov_test):

{
  "<scene_id>": {
    "pred_classes": (N_instances,)            int,   # 1=rotation, 2=translation
    "pred_scores":  (N_instances,)             float, # confidence
    "pred_masks":   (N_points, N_instances),          # >1e-3 = point in instance
    "pred_origins": (N_instances, 3)           float, # articulation axis origin
    "pred_axises":  (N_instances, 3)           float, # articulation axis direction
  },
  ...
}

Interactable phases (inter_dev, inter_test): same, minus pred_origins/pred_axises — no articulation to predict. Note: pred_classes here is still the rotation/translation class of the movable object each interactable part belongs to, not a separate class.

N_points must match the input point cloud exactly (row-for-row, same order — see §2 on using the provided preprocessed data to guarantee this). N_instances is up to you.

If you're using the USDNet baseline: general.debug=true during inference (see scripts/infer_mov.sh/infer_inter.sh) dumps debug/val_preds/preds.pkl already in this format — submit that file directly. Don't run it through transfer_preds_files.py or the zip-based steps at the end of those scripts — that repackages predictions into a different (.txt-based) format for a different evaluation convention, and this server only accepts .pkl.

Rank	Team / Method	MAP_AXIS_ORIGIN ↑
🥇	pico-mr Winners Challenge @ICCV25	0.41
2	astar	0.29
3	Uni Stuttgart	0.27
4	ww	0.26
—	USDNet Official Baseline	0.26

Articulate3D ChallengeOpen

What is the task?

📊 Evaluation

🎯 Challenge Timeline

📄 Paper Track Timeline

Working with Articulate3D

📦 Articulate3D Annotations

Dataset Structure

Part Segmentation

Articulation Annotations

🗃️ Data Loader