TianhangCheng7/IMU4DData

IMU4D Data

Processed motion / IMU training and evaluation data for IMU4D. The repo mirrors the data/processed/ tree of the IMU4D_dev code base, so downloading it into $IMU4D_DATA_ROOT/processed reproduces every default path used by the training configs.

Dataset	Path	train / val / test	Size
MotionMillion + LINGO	motionmillion/v1/wds/	913,308 / 58,444 / 170,757	~111 GB
HiPHI	hiphi/v1/{wds,splits}/	13,037 / 500 / 500	~13 GB
OMOMO	omomo/v1/{wds,splits,samples}/	5,279 / 500 / 500	~1.1 GB
HUMOTO	humoto/v1/{wds,splits,samples}/	685 / 50 / 50	~0.3 GB
Qwen3-0.6B caption rewrites	<dataset>/qwen3_0.6B_rewrite_v2/<split>/*.jsonl	HiPHI, HUMOTO, OMOMO: all local sidecars; MotionMillion: test only (23 files)	~230 MB
OMOMO canonical object meshes (upright, bbox-aligned, front = +z; 2026-09-15 frames, see dataset_process/asset_canonical_frames.json in IMU4D_dev)	omomo/canonical_objects/<asset>.obj	15 assets + 2 part meshes	~24 MB
HiPHI canonical object meshes (same convention; every shipped mesh incl. its __mirror twin; 2026-09-16: mirror frames recomputed on the mirrored geometry)	hiphi/canonical_objects/<asset>.obj	40 assets + 40 mirrors	~98 MB
HUMOTO canonical object meshes (same convention)	humoto/canonical_objects/<asset>.obj	66 assets	~132 MB
IMUPoser (real-world, measured IMUs)	imuposer/v2/{wds,splits,samples}/	133 / 34 / 34	~0.13 GB
NCSA phone / watch / earbud (real-world, GVHMR pseudo-GT, joint-angle-only supervision)	ncsa/v3/{wds,splits,samples}/	27 / 6 / 6 sessions	~66 MB
NCSA meeting room, 3-point multi-view (2026-09-15 calibrated capture)	ncsa/meeting_room_3pt_mv/{wds,splits,samples,metadata}/	32 / 15 / 15 clips	~51 MB
NCSA meeting room, 2-point multi-view (same capture protocol, headphone dropped)	ncsa/meeting_room_2pt_mv/{wds,splits,samples,metadata}/	16 / 6 / 6 clips	~24 MB

Real-world variants (details in dataset_process/realworld/README.md and dataset_process/ncsa/README_ncsa_imu.md of IMU4D_dev):

• imuposer/v2: world frame rotated z-up -> y-up, gravity-free acceleration (imu_acc_add_gravity = False). The uncorrected first conversion (v1) is not hosted any more; rebuild it from the legacy per-sequence pickles (data/raw/imuposer/legacy_v1, not hosted) with the plain converter if you need it.
• ncsa/v3 (2026-09-14 build, supersedes the v1 / v2 variants, which are no longer hosted): the 34-session 30 Hz capture (phone on the right thigh, watch on the left wrist, earbud on the right ear; slots 1 / 4 / 3, earbud copied to slot 2, slots 0 and 5 empty) with acceleration sign fix and gravity-free acceleration, and GVHMR static-camera labels instead of PromptHMR (convert_ncsa_imu.py --motion-source gvhmr --joint-angle-only --acc-fix --no-loader-gravity). Because the monocular labels only pin down local joint angles, the samples carry a motion_supervise mask that hides root orientation and translation from the loss — train on this set, but do not read its global trajectory as ground truth. Session scene_2_seq_2 is excluded (uncorrelated IMU orientations). 27 train / 6 val sessions, 45,943 / 8,016 frames; val and test are the same six sessions.
• ncsa/meeting_room_3pt_mv and ncsa/meeting_room_2pt_mv (2026-09-15 multi-camera capture in a calibrated meeting room, labels fused across views): clips of ids ncsa/meeting_room/0915_scene_<s>/seq_<q>/c<NN>, one sample per session and camera. 3pt_mv keeps all three devices (10 sessions, 32 train / 15 test clips, 36,756 / 5,052 frames), 2pt_mv drops the headphone (5 sessions, 16 train / 6 test clips, 17,154 / 2,031 frames). Train and test are a temporal split inside each session with a 2 s guard band, so they never share frames. metadata/blacklist_v1.json lists the manually flagged label intervals; they are already removed at pack time (clean_ranges in splits/summary.json) and are shipped only as a record. val is an alias of test in the manifests. These two sets are the held-out metric for real-IMU fine-tuning; the id namespace does not overlap v3, so there is no leak.

Format

Every WDS root contains wds/manifest.json and wds/{train,val,test}/*.tar (WebDataset). Each sample is a pickle with motion_data_smpl85 (T, 85), imu_traj (T, 6, 6), texts, source, id; HiPHI / OMOMO / HUMOTO samples also carry object poses and identities. Real-world samples (IMUPoser, NCSA) store measured imu_acc (T, 6, 3) and imu_ori (T, 6, 3, 3) in model sensor order instead of imu_traj, plus imu_acc_add_gravity / imu_missing_slots flags read by the loader. Caption rewrites are JSONL sidecars keyed by sample id, one file per shard.

Only v2 caption rewrites are hosted. MotionMillion v2 currently covers test shards only; no NCSA text annotations or v2_frac10 / v2_frac50 subsets are hosted.

Download

pip install "huggingface_hub[cli]" webdataset
export IMU4D_DATA_ROOT=/path/to/IMU4D_dev/data

# everything (~130 GB)
hf download TianhangCheng7/IMU4DData --repo-type dataset \
  --local-dir "$IMU4D_DATA_ROOT/processed"

# a single dataset
hf download TianhangCheng7/IMU4DData --repo-type dataset \
  --local-dir "$IMU4D_DATA_ROOT/processed" --include "omomo/**"

# real-world fine-tune sets only (<0.5 GB)
hf download TianhangCheng7/IMU4DData --repo-type dataset \
  --local-dir "$IMU4D_DATA_ROOT/processed" --include "imuposer/**" --include "ncsa/**"

See the IMU4D_dev README (Datasets section) for how training and evaluation stream these shards and for the environment variables that override each root.

Not included

The original HiPHI / OMOMO / HUMOTO releases (data/raw/) are not redistributed; obtain them from the original authors. Only the per-asset canonical object meshes (rotated copies of the release object meshes, */canonical_objects/) are hosted, for the object-track re-expression in the IMU4D_dev loader. The legacy HuMoTo pickles (raw/humoto/v1/humoto_data/{all,all_time,*_indices.npy}, the rebuild source of humoto/v1) and the legacy IMUPoser / DIP-IMU per-sequence pickles (raw/{imuposer,dipimu}/legacy_v1, rebuild sources of the v2 sets) are part of data/raw/ and are not hosted either. DIP-IMU is not included either (its license does not permit redistribution): project members with a DIP-IMU license fetch dipimu/v2 from the private companion repo TianhangCheng7/IMU4DData_private, everyone else builds processed/dipimu/v2 from the official DIP_IMU_processed.zip as the IMU4D_dev README explains (dataset_process/realworld/build_corrected_variants.sh dipimu). The NCSA raw capture (data/raw/ncsa/: videos, monocular/multi-view body fits) is not included either, only the converted IMU + SMPL-X clips. Intermediate build products (hiphi/v1/{smpl85,intermediate,samples}, previews, Rerun recordings) are omitted. The SMPL-X body model must be downloaded separately from https://smpl-x.is.tue.mpg.de under its own license.

Sources: MotionMillion, LINGO, HiPHI, OMOMO, HUMOTO, IMUPoser, DIP-IMU (not redistributed), NCSA phone / watch / earbud capture (UIUC NCSA; 2026-08-30 + 2026-09-13 handoffs with GVHMR pseudo ground truth, 2026-09-15 calibrated multi-view meeting-room capture). Please follow each source dataset's license and citation requirements.