JimSYXu/SAM

Dataset Overview

• Name: SAM (Spatial Audio-driven human Motion)
• Purpose: This dataset provides paired human motion sequences and corresponding spatial audio for modeling and analysis.
• Source: Captured using a Vicon motion capture system.
• Size: Over 9 hours of recordings at 120 frames per second (FPS).

Updates

Added audio.txt reference file. This file maps audio IDs to their corresponding class labels. The data is segmented by the human subject who performed the associated motions.

Folder Structure

SAM/
|- audio/
│- bvh/
|- c3d/
|- motion/
|- retarget/
|- splits/
|- stereo_audio/
|- ignore_list.txt
|- ssl.npy
|- subject_ids.txt

• audio/: Stereo audio clips that have already been separated into left and right channels.
• bvh/: Raw BVH motion-capture exports from the Vicon Shogun pipeline.
• c3d/: Raw C3D recordings from the same capture sessions, preserved for custom processing.
• motion/: MOSPA-ready motion files in NPZ format, including the sound-source location in the character’s local space.
• retarget/: BVH files after the retargeting step, prior to conversion into the NPZ motion assets.
• splits/: Dataset split definitions (train/val/test) stored as three TXT files indexed by IDs.
• stereo_audio/: Stereo audio sequences aligned with each motion sample via shared IDs.

Important Files

• ignore_list.txt: IDs flagged for questionable motion or audio; skip them at your discretion.
• ssl.npy: Global sound-source coordinates (AMASS frame). Shape (12, 4, 3) = 12 subjects × 4 speakers × XYZ.
• subject_ids.txt: Maps every motion sequence ID to its subject.

How to Use This Dataset

Usage and baseline pipelines: MOSPA.

ID Encoding

Each sample ID is a 7-digit token:

Digits	Meaning	Values
0–2	Audio clip
3	Source direction	0 front, 1 right, 2 left, 3 back
4	Speaker	0 Speaker_A_R, 1 Speaker_A_L, 2 Speaker_B_R, 3 Speaker_B_L
5	Distance	0 close, 1 far
6	Motion genre	0 insensitive, 1 neutral, 2 sensitive

Known Issues

We recommend retargeting motions yourself; raw BVH/C3D files are provided for that purpose. Some sound-source annotations contain mocap noise, so consider using the fixed global locations defined in ssl.npy.