WJM-George/063_erehps

Sphere360 processed

Packed export of a local Sphere360 V2A / VT2A workspace: 4-channel FOA audio, the most-dynamic continuous 10 second window per clip, VideoMAE-v2 ViT-B features, and VAE latents.

This repository does not include the original 360° videos.

Repo: WJM-George/063_erehps

Layout

Path	What
parquet/full_{train,test}.parquet	Full-length FOA clip index + tar pointers
parquet/dynamic10_{train,test}.parquet	10s windows + VideoMAE + latent + tar pointers
audio_foa/{split}/shard-XXXXXX.tar	Uncompressed tar shards of full-length 4ch WAV
audio_foa_dynamic10/{split}/shard-XXXXXX.tar	Uncompressed tar shards of 10s 4ch WAV
manifests/	Per-collection shard inventories

Counts: 30,000 train / 1,000 test full-length clips; 29,987 train / 1,000 test dynamic-10s windows. Audio is 4ch FOA, 44.1 kHz, pcm_f32le. Latents are float32 with shape (64, 431). VideoMAE-v2 features are float32 with shape (T, 768) where T is usually 5.

Read a row and its WAV

import io, tarfile
import numpy as np
import pyarrow.parquet as pq
import soundfile as sf
from huggingface_hub import hf_hub_download

repo = "WJM-George/063_erehps"
table = pq.read_table(hf_hub_download(repo, "parquet/dynamic10_train.parquet", repo_type="dataset"))
row = table.slice(0, 1).to_pydict()
i = 0
latent = np.frombuffer(row["latent"][i], dtype=np.float32).reshape(row["latent_shape"][i])
videomae = np.frombuffer(row["videomae"][i], dtype=np.float32).reshape(row["videomae_shape"][i])

tar_path = hf_hub_download(repo, row["audio_shard"][i], repo_type="dataset")
with tarfile.open(tar_path) as tar:
    wav_bytes = tar.extractfile(row["audio_member"][i]).read()
audio, sr = sf.read(io.BytesIO(wav_bytes))

Notes

• Shards are uncompressed tars (~4 GiB). Do not gzip them; the WAV payload is already PCM.
• Absolute /mnt/... paths from the original workspace were stripped.
• dynamic_usable is the original window-quality flag; some 10s windows are padded at the end of a short clip.