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. We use this refined dataset for 3D world generation and reconstruction in our Puffin-World.

Sample Videos

Each clip is a 1×3 comparison — RGB | Original Depth | Our Aligned Depth — with depth rendered by vision banana representation. It shows how the DA3-aligned depth (right) densifies and cleans up the original sparse depth (middle).

Directory structure

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

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: DL3DV-Absolute-Camera.

Caption Pipeline

Beyond the aligned dataset, we also release a complete captioning pipeline for annotating the dense depth map for arbitrary datasets, aligning with the sparse depth, and visualizing the corresponding depth maps. The pipeline is available in our GitHub repository.