PRISM Dataset

PRISM (Pressure and Inertial Sensing for Human Motion and Interaction) is a multimodal motion-capture dataset combining SMPL body, 8-IMU body sensors, insole pressure, SLAM head trajectory + environment point cloud, and scene objects. It covers everyday locomotion, balance, sports, and stepping/sitting interactions in light and dark conditions.

Released with GRIP: Ground Reaction Inertial Poser (CVPR 2026), which fuses 4 wrist/foot IMUs with insole pressure (GRF, CoP, contact) to reconstruct physically plausible full-body motion.

Dataset access is gated by a short Google Form. Submit the form and the download link will appear on the response screen.

After downloading, unzip under data/:

unzip PRISM.zip -d data/

Trials: Walking, Jogging, Switch Direction, Sidestep, Lunge Walk, Forward Jump, Squatting, Stretching, Balancing, Object Carrying, Stepping Boxes/Stair, Sitting Boxes, Soccer, Tennis, Baseball, Golf. Dark = low-light recording. Stepping* / Sitting* include scene-object meshes.

PRISM/
├── data/PRISM/<subj>/<take>.pkl   # one full-length take per file
├── json/
│   ├── dataset_split.json         # sequence-level train / test split
│   └── dataset_statistics.json    # per-sequence chunk windows + flags
├── image/PRISM.png
├── data_viewer.py                 # aitviewer-based visualization
├── requirements.txt
└── README.md

Each .pkl holds a full-length recording. The JSON files in json/ define how each take is sliced into 1000-frame (10 s) chunks (seq000 = frames 0–1000, seq001 = 1000–2000, …) and assigns each chunk to train or test.

Each .pkl is a Python dict. F is the take's frame count (e.g. 13,160).

subj_info : {subj_id, gender, age, height, weight, shoe_length, arm_length}
data_info : {take_id, trial_name, insole_size, fps,
             synth_imu_frames: {L_Wrist: [...], R_Wrist: [...]}}

synth_imu_frames lists frame indices where the wrist IMU was reconstructed from SMPL because the Apple Watch dropped or returned corrupted samples. Test chunks are guaranteed to contain none.

Parts: L_Foot, R_Foot, L_Wrist, R_Wrist, Head, Pelvis, L_Knee, R_Knee. Field names encode <quantity>_<frame>_<processing>.

Per foot (L_Foot, R_Foot):

combined.{force_world, CoP_world} provide both-feet sums.

Present only for Stepping* / Sitting* trials, otherwise None. Each entry: {vertices [V,3], faces [Fc,3], centroid [3], extents [3]}.

• World frame is Z-up.
• World translation = smpl_params['trans'] + smpl_params['root_offset'].
• IMU mounting positions (approximate SMPL vertex/joint indices) are defined as IMU_VERTEX_IDX / IMU_JOINT_IDX in data_viewer.py.

json/dataset_split.json:

{
  "train": ["subj001_take002_seq000", ...],
  "test":  ["subj001_take002_seq004", ...]
}

Chunk-level random split (seed = 42); the same take can contribute chunks to both splits. Chunks containing any synthesized wrist-IMU frames are kept on the train side only, so the test set is fully measured.

json/dataset_statistics.json contains per-sequence records:

{subject_id, take_id, sequence_id, chunk_idx,
 start_frame, end_frame, total_frames, has_synthetic_imu, file_name}

Slice the corresponding .pkl with start_frame:end_frame.

conda create -n prism python=3.11 -y && conda activate prism
pip install --no-build-isolation -r requirements.txt

Tested on Python 3.11 with NumPy 1.26.x. --no-build-isolation is required by chumpy's setup.py.

Load a take:

import pickle

with open("data/PRISM/subj001/take002.pkl", "rb") as f:
    data = pickle.load(f)

poses    = data["smpl_params"]["poses"]                # [F, 72]
foot_acc = data["imu"]["L_Foot"]["acc_world_filt"]     # [F, 3]
contacts = data["insole"]["L_Foot"]["contacts"]        # [F, 2]
head     = data["slam"]["head_traj"]                   # [F, 3]

Iterate the split:

import json, pickle

split = json.load(open("json/dataset_split.json"))
stats = json.load(open("json/dataset_statistics.json"))
meta  = {f"{r['subject_id']}_{r['take_id']}_{r['sequence_id']}": r
         for r in stats["train_sequences"] + stats["test_sequences"]}

for seq_id in split["train"]:
    r = meta[seq_id]
    take = pickle.load(open(f"data/PRISM/{r['subject_id']}/{r['file_name']}", "rb"))
    chunk = take["smpl_params"]["poses"][r["start_frame"]:r["end_frame"]]  # [1000, 72]

Interactive visualization — overlays SMPL, IMU, GRF, foot contact, head trajectory, point cloud, and object mesh:

python data_viewer.py                          # random order, infinite loop
python data_viewer.py -s subj001               # all takes from subj001 (one pass)
python data_viewer.py -s subj001 -t take002    # one specific take
python data_viewer.py --data-dir other/PRISM   # use a different data root