README.md

---
license: apache-2.0
task_categories:
- depth-estimation
tags:
- depth
- depth-anything-3
- dl3dv
- metric-depth
size_categories:
- 1M<n<10M
---

# DL3DV-Depth-DA3-Aligned

Per-frame **depth** annotations for the **DL3DV** dataset, produced by
**Depth-Anything-3 (DA3)** and then **aligned** to each scene's sparse depth
from the original DL3DV reconstruction.

## Directory structure

The archive mirrors the source DL3DV / DL3DV-ALL-960P layout — **one `.zip` per
scene**, grouped into bucket folders `1K`–`7K` (`<bucket>/<scene_hash>.zip`).
The scene hashes match DL3DV-ALL-960P and `KangLiao/DL3DV-Absolute-Camera`, so
depth pairs 1:1 with the source frames / camera annotations.

Each `<scene_hash>.zip` unpacks to:

```
dense/
└── depth_da3/
    ├── frame_00001.npy
    ├── frame_00002.npy
    ├── frame_00003.npy
    └── ...
```

Each `frame_NNNNN.npy` is a **float32** depth map — `np.load(...)` returns an
array of shape `(H, W)` (e.g. `(536, 954)`), one per source frame, indices
matching the DL3DV frames.

## How the depth was produced

- **Predicted** with **Depth-Anything-3** (DA3).
- **Aligned** to the **sparse depth** of the original DL3DV dataset (per-scene
  alignment against the sparse reconstruction), so each scene's DA3 depth is
  brought into a consistent, scale-aligned space.

## Usage

```python
import numpy as np
depth = np.load("dense/depth_da3/frame_00001.npy")  # (H, W) float32
```

## Notes

- ~6,377 scenes; each `.npy` frame ≈ 2 MB (float32), stored losslessly.
- Companion camera annotations: `KangLiao/DL3DV-Absolute-Camera`.