---
license: cc-by-4.0
task_categories:
- video-classification
size_categories:
- 100K<n<1M
tags:
- video
- video-compression
source_datasets:
- nkp37/OpenVid-1M
---

# OpenVid GOP cache for HFTV

Fixed-size one-second video clips, packed for training a spatiotemporal
autoencoder. Derived from [OpenVid-1M](https://huggingface.co/datasets/nkp37/OpenVid-1M)
(Nan et al., *OpenVid-1M: A Large-Scale High-Quality Dataset for
Text-to-video Generation*, ICLR 2025), used and redistributed under
**CC BY-4.0**; this derivative carries the same licence.

## Why it exists

OpenVid-1M ships as 12.4 TB of ~50 GB zip parts, so selecting clips by
caption cannot reduce what you *fetch* -- only what you keep. This is
that selection done once: the filter below applied, every clip decoded
to a fixed geometry, and the result packed into a flat memmap that a
training job reads directly. Built from 20 of 186 parts.

## Contents

| directory | clips | GOPs | filter |
|---|---:|---:|---|
| `openvid-head/` | 10544 | 31051 | static, low motion, >= 1 s, **and** a talking-head caption |
| `openvid-diverse/` | 28991 | 72831 | static, >= 1 s, quota-balanced over nine caption categories |

`openvid-diverse/` is quota-balanced across animal, food, vehicle,
text-graphic, urban, nature, interior, object and people-other, each
assigned by first match in priority order so every clip has exactly one
label. 85% of each quota is low motion (score < 5) and 15% is
5-15, deliberately: a corpus with no motion in it is its own kind of
narrow.

The balance is computed over the clips **actually reachable in the
fetched parts**, not corpus-wide, because a build fetches a handful of
~45 GB parts and a corpus-wide quota would leave the published balance
as whatever those parts happened to hold. That is how an earlier
version of this dataset ended up with a "non-person" split that was
almost entirely stock landscape footage.

Talking-head captions match:

```
speaking
interview
talking
microphone
looking at the camera
portrait
podcast
news anchor
```

## Format

Each directory holds `gops.u8` and `index.json`.

- `gops.u8` -- raw `uint8`, C-order, shape `(count, 12, 120, 160, 3)`,
  RGB. One row is one GOP: **12 frames at 12 fps = 1.000 s**.
- `index.json` -- `{frames, height, width, fps, count, sources}`, where
  `sources[i]` is the OpenVid filename GOP `i` came from. Up to
  3 consecutive GOPs come from one clip.

```python
import json, numpy as np
meta = json.load(open("openvid-head/index.json"))
gops = np.memmap("openvid-head/gops.u8", np.uint8, "r",
                 shape=(meta["count"], 12, 120, 160, 3))
```

Clips are **centre-cropped to 4:3 and then scaled** -- never stretched.
Letterbox bars are detected and removed before the crop.

## Caveats

- 120x160 at 12 fps is a deliberate reduction, chosen for a
  narrowband analog video mode. This is not a general-purpose video
  corpus; nothing here is at OpenVid's original quality.
- The filter is metadata-driven. `camera motion` and `motion score`
  are OpenVid's own annotations and inherit their errors.
- Captions are matched by regular expression, so both splits contain
  what a caption *says*, not what a detector found. The non-person
  split in particular is "no person was mentioned", which is not the
  same as "no person is visible" -- expect some contamination, and
  check the pictures rather than trusting the label.

Built by [`scripts/prepare_openvid_job.py`](https://github.com/arodland/HFTV).