README.md

---
license: other
license_name: diogenes-evaluation-only
license_link: LICENSE
language:
  - en
tags:
  - video
  - depth
  - depth-estimation
  - camera-pose
  - 3d-reconstruction
  - world-model
  - gameplay
  - non-trainable
size_categories:
  - n<1K
extra_gated_heading: Request access to Diogenes Depth + Pose Sample v01
extra_gated_description: >-
  Access is restricted. Enter the access code supplied to you by Diogenes Lab.
  Requests submitted without a valid code are rejected. This dataset must not be
  used to train, fine-tune, or evaluate any model.
extra_gated_button_content: Request access
extra_gated_fields:
  Access code: text
  Organization: text
  Intended use: text
  I confirm I will NOT use this data to train, fine-tune, or evaluate any model: checkbox
---

# Diogenes Depth + Pose Sample v01

**A sample dataset.** One complete 120-second capture package: gameplay video, a
per-frame linear depth map, and a per-frame camera pose — all on a single clock, so
every pixel of depth can be back-projected into one shared world frame.

This is the artifact behind the claim that our capture is *geometric*, not just video.
It exists so an evaluator can reproduce the reconstruction themselves rather than take
our word for it.

## ⚠️ Not for model training

Every package in this repository is `non_trainable` and `production_certified: false`.
The source runs in a release-test capture mode and carries no consent grant.

Permitted: inspecting capture quality, validating the depth/pose schema, verifying
time alignment, and reproducing the back-projection. **Not permitted:** training,
fine-tuning, or evaluating any model; redistribution outside approved collaborators.

The video depicts **Fallout 4**, copyright Bethesda Softworks LLC. Diogenes Lab claims
no rights to the depicted game content.

## What is in the package

`data/2026_08_06_FO4Commonwealth_120s_01/`

| File | Size | Content |
|---|---:|---|
| `video.mp4` | 239.8 MB | 1920×1080, 30 fps CFR, H.264, 3600 frames = 120.000 s, in-game audio |
| `depth.zip` | 364.0 MB | 1800 × `depth_<even frame>.exr` — 1280×720, fp16, channel `Z` |
| `action_camera_30fps.json` | 3.2 MB | 3600 rows, one per video frame |
| `action_camera.json` | 1.7 MB | 1800 rows at 15 fps, the depth-frame subset |
| `systeminfo.json` | 110 B | Capture surface geometry |
| `sha256.json` | 1.1 KB | Digest for all 11 delivered files |
| `QC/` | 41.8 MB | Conversion manifest, frame-pair report, overlay preview, source-frame selection |

`preview/spatial-overview-en.mp4` is a rendered walkthrough of this exact package —
RGB with an action HUD, the depth map in metres, and the fused point cloud with the
camera path. Nothing in it is staged; it is rendered from the files above.

## Per-frame pose row

```json
{
  "time": "...", "fps": 30.0, "frame": 0,
  "camera_position": [x, y, z],
  "camera_rotation_quaternion": [x, y, z, w],
  "camera_intrinsics": {"fx": 698.2612, "fy": 698.26117, "cx": 960.0, "cy": 540.0},
  "camera_follow_offset": [...], "camera_speed": [...],
  "player_position": [...], "player_rotation_quaternion": [...], "player_speed": [...],
  "mouse_dx": 0, "mouse_dy": 0, "keyCode": [], "metric_scale": 1.0
}
```

## Reading the depth — three conventions you need

These are not guessable from the files; they were established empirically and are stated
here so the reconstruction is reproducible on first attempt.

1. **The stored quaternion is world→camera.** To go the other way, transpose:
   `R_cam2world = R(q)ᵀ` (equivalently, take the conjugate).
2. **The depth EXR is 1280×720 while the intrinsics are calibrated for 1920×1080.**
   Scale the intrinsics by 2/3 before using them with the depth raster.
3. **First-person viewmodel geometry sits at ≈0.21–1.0 m** with a clean empty gap
   between it and the scene. Discard `z ≤ 1.05` to drop the weapon/armature.

Back-projection, left-handed Y-up, image `v` downward:

```
p_cam   = [(u - cx)/fx * z,  -(v - cy)/fy * z,  z]
p_world = R(q)ᵀ · p_cam + camera_position
```

Depth is **view-space linear Z in metres**. `NaN` marks sky and invalid samples.

## Measured properties of this package

| Property | Value |
|---|---|
| Duration | 120.000 s, 3600 video frames, 1800 depth frames |
| Depth-to-video pairing coverage | 1.000 (1800 / 1800 eligible even frames) |
| Frame-sync watermark mismatches | 0 of 3600 decoded |
| Camera path length | 324.80 m |
| Elevation range | 7.38 m |
| Cross-frame depth consistency | 0.15 % median-of-medians reprojection error over 12 frame pairs spaced 10 frames apart |

The consistency figure is the honest test: back-project frame *A*, reproject it into
frame *B* using only the stored poses, and compare against *B*'s own depth. A broken
handedness or intrinsic scale puts this in the 7–9 % range; 0.15 % means the pose and
depth agree.

## Known limitations

Stated up front rather than discovered later:

- `fp16` with two-decimal truncation cannot represent 1 cm beyond about `z ≥ 16 m`, so
  distant surfaces show stepped quantisation.
- The depth raster includes the first-person viewmodel (the ≈0.21 m block). It is real
  content of the EXR and is not masked out in the delivered file.
- 36 reconstructed frames carry a ~33 ms micro-jump from dropped-frame reconstruction.
- Dynamic objects (NPCs, animals) leave ghost trails in any multi-frame fusion; only the
  player and the viewmodel are explicitly removable.

## Related

- `DiogenesLab/Diogenes_Gameplay_raw_sample_v01` — gameplay video with frame-aligned
  keyboard/mouse action annotations.
- `DiogenesLab/Diogenes_Gameplay_validation_sample_v01` — the same capture line packaged
  as fixed-length shards.

Access to all Diogenes Lab sample datasets uses the same access code, but permission is
granted per repository — request each one separately.