BumsooKim00/nisr-dataset

NISR Dataset — Neural Inverse Sound Rendering

Dataset for training a model that predicts natural frequencies and 3D mode shapes from an impact sound recording and a 3D mesh.

---

Changelog

Version	Date	Objects	Description
v1.0	2026-01	100	Initial release — ObjectFolder-Real (100 everyday objects)
v2.0	2026-06-20	1100	Extended with ObjectFolder 2.0 (+1000 objects, obj_id 101–1100)
v3.0	2026-06-21	1100	Added physics-simulation-based impact sounds (sim_*.wav, 1092 objects) — hit points extracted from a PyBullet drop simulation instead of a fixed index; 8 objects skipped (degenerate contact geometry)
v4.0	2026-06-25	1100	Added randomized single-hit variations (rand_{mat}_{i}.wav, 8 per material) — pure modal re-synthesis (no physics sim) with randomized hit time, hit point, and hit strength; hit params stored in rand_params.npz

Current version: v4.0 (2026-06-25) obj_id 1–100: ObjectFolder-Real · obj_id 101–1100: ObjectFolder 2.0

---

Derived from ObjectFolder-Real (100 everyday objects) and ObjectFolder 2.0 (1000 objects). Modal data generated via FEM (LOBPCG) simulation with 8 material configurations per object.

---

Task

Input:  impact sound wav  +  3D mesh (voxel 32³)  +  object size L  +  material (E, ρ, ν)
Output: natural frequencies (k,)  +  3D mode shapes (B, 3, k)

---

Dataset Statistics

Objects	1100 (ObjectFolder-Real 100 + ObjectFolder 2.0 1000)
Materials per object	8
Modes per sample	up to 20
Audio	2.0 sec @ 44,100 Hz
Total samples	8800 (1100 objects × 8 materials)
Total size	~22 GB

---

Structure

training_dataset/{obj_id}/
├── voxel.npz               ← 3D voxel grid (32³), shared across materials
├── feat/feat_{mat}.npz     ← training labels (freqs, mode shapes)
└── wav/
    ├── sound_{mat}.wav        ← impact sound — fixed hit point (v2)
    ├── sim_{mat}.wav          ← impact sound — physics-simulation hit points (v3)
    ├── rand_{mat}_{i}.wav     ← impact sound — randomized single hit, i = 0..7 (v4)
    └── rand_params.npz        ← hit params (t_start, hit_index, gain) per variation (v4)

Sound variants

File	Hit point	Description
sound_{mat}.wav	fixed (index 0)	Single impact at a fixed boundary voxel. Available for all 1100 objects.
sim_{mat}.wav	physics-derived	Object is dropped in a PyBullet simulation; the contact points and impact timing from the drop drive a multi-contact, force-weighted modal synthesis — producing realistic multi-bounce impact sounds. Available for 1092 objects (see below).
rand_{mat}_{i}.wav	randomized (8×)	Eight single-hit variations per material, each with a randomized hit time, hit point (boundary voxel), and hit strength (gain). Pure modal re-synthesis — no physics simulation. Available for all 1100 objects. Per-variation params in rand_params.npz.

rand_params.npz (v4)

Per-object hit parameters for the randomized variations:

Key	Shape	Description
materials	(8,)	material names, row order of the arrays below
t_start	(8, 8)	impact onset time (sec), range [0.10, 0.70]
hit_index	(8, 8)	boundary voxel index of the hit
gain	(8, 8)	hit strength → loudness scale, range [0.40, 1.00]

Array shape is (n_materials=8, n_variations=8); index [m, i] corresponds to rand_{materials[m]}_{i}.wav. Variations are deterministic (RNG seeded by obj_id).

Coverage of sim_*.wav (v3)

Physics-simulation sounds were generated for 1092 / 1100 objects. The following 8 objects are skipped because their mesh geometry is degenerate for rigid-body contact:

950, 1001, 1006, 1008, 1016, 1017, 1018, 1020

These meshes are either extremely flat or abnormally scaled (e.g. obj 1001 is ~2.9 units across). After the drop, the rigid body settles deeply embedded in or flush against the ground plane, and PyBullet's convex-hull getContactPoints fails to build a contact manifold — returning 0 contacts regardless of timestep, solver iterations, maximal-coordinates, or collision-margin tuning. (Geometric contact clearly exists — getClosestPoints returns 4000+ pairs — but no force-bearing manifold is formed, so no impact event can be extracted.)

For these 8 objects, only the fixed-hit-point sound_*.wav is provided.

---

Splits

Train/val/test split is not yet defined. All samples are currently unsplit. Split assignment will be added in a future release.

---

Download

Option 1 — Python (recommended)

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="BumsooKim00/nisr-dataset",
    repo_type="dataset",
    local_dir="./nisr-dataset",
)

Option 2 — git clone

git lfs install
git clone https://huggingface.co/datasets/BumsooKim00/nisr-dataset

Option 3 — wget (single file)

# voxel
wget https://huggingface.co/datasets/BumsooKim00/nisr-dataset/resolve/main/training_dataset/1/voxel.npz

# feat
wget https://huggingface.co/datasets/BumsooKim00/nisr-dataset/resolve/main/training_dataset/1/feat/feat_Ceramic.npz

# wav
wget https://huggingface.co/datasets/BumsooKim00/nisr-dataset/resolve/main/training_dataset/1/wav/sound_Ceramic.wav

No login required (public repo). Python: pip install huggingface_hub

---

Loading

import numpy as np
import scipy.io.wavfile as wav

obj_id = 1
mat    = "Ceramic"
base   = f"training_dataset/{obj_id}"

# audio input — fixed hit point (v2)
fs, audio = wav.read(f"{base}/wav/sound_{mat}.wav")
audio = audio.astype(np.float32) / 32767.0   # int16 → float32

# audio input — physics-simulation hit points (v3, recommended)
fs, audio_sim = wav.read(f"{base}/wav/sim_{mat}.wav")
audio_sim = audio_sim.astype(np.float32) / 32767.0

# audio input — randomized single-hit variations (v4)
i = 0
fs, audio_rand = wav.read(f"{base}/wav/rand_{mat}_{i}.wav")
audio_rand = audio_rand.astype(np.float32) / 32767.0
params = np.load(f"{base}/wav/rand_params.npz")
mats   = list(params["materials"])
m      = mats.index(mat)
t_start, hit_index, gain = (params["t_start"][m, i],
                            params["hit_index"][m, i],
                            params["gain"][m, i])

# mesh input
voxel = np.load(f"{base}/voxel.npz")["voxel"]       # (32, 32, 32)

# labels
feat     = np.load(f"{base}/feat/feat_{mat}.npz")
freqs    = feat["freqs"]      # (k,)      natural frequencies in Hz
feats_in = feat["feats_in"]   # (B, 3, k) mode shapes on boundary voxels
coords   = feat["coords"]     # (B, 3)    boundary voxel coordinates
surface  = feat["surface"]    # (B, 6)    surface normal encoding

---

Materials

Material	ρ (kg/m³)	E (Pa)	ν
Ceramic	2700	7.2e10	0.19
Glass	2600	6.2e10	0.20
Wood	750	1.1e10	0.25
Plastic	1070	1.4e9	0.35
Iron	8000	2.1e11	0.28
Polycarbonate	1190	2.4e9	0.37
Steel	7850	2.0e11	0.29
Tin	7265	5.0e10	0.325

---

Generation

FEM modal analysis (LOBPCG) → modal feature extraction → modal sound synthesis.

See GENERATE.md for the full pipeline.

---

License

CC BY 4.0 Mesh data from ObjectFolder-Real — original license applies to source meshes.