Delivery format

HDF5 & zarr delivery format for egocentric episodes

One group per episode, time-major chunks, per-stream nanosecond timestamps.

By the Firsthand capture teamLast updated September 23, 2026

Short answer

HDF5 and zarr store one group per episode with time-major chunking and per-stream nanosecond timestamps. This is the format Firsthand’s own QA tooling reads, so it is the most complete — ideal for random access into long episodes and for verifying sync yourself.

Field mapping

How Firsthand fields map into this format

Container: One .h5 or .zarr group per episode
Chunking: Time-major, 30-frame chunks, blosc-zstd level 5
Timestamps: int64 nanoseconds per stream, PTP domain
Groups: streams/, annotations/, calibration/ mirroring the episode schema
Use case: Random access into long episodes, sync verification

Loading it

load_hdf5.py

import h5py, numpy as np

with h5py.File("data/EP_0117_KITCHEN_A.h5", "r") as f:
    rgb = f["streams/head_rgb"][100:130]        # a 30-frame chunk
    depth = f["streams/depth_mm"][:]
    hands = f["annotations/hand_joints"][:]      # [T, 2, 21, 3]
    ts_head = f["streams/head_rgb"].attrs["timestamp_ns"]

Gotchas

What to watch for

Read whole chunks (30 frames) rather than single frames to avoid decompression overhead.
zarr is the better choice for object storage; HDF5 for a single local file.
The calibration/ group carries both the opening and closing solves for drift checks.

Get a HDF5 & zarr sample.

40 episodes across 4 environments, delivered in the format your stack already reads.

Get sample data →All formats