androidli's picture
Add model/dataset card (README.md)
d343d7b verified
|
Raw History Blame Contribute Delete
3.09 kB
---
task_categories:
- depth-estimation
- image-to-image
tags:
- depth
- pseudo-labels
- monocular-depth
- depth-anything-v2
- high-resolution
---
# hires1920 Depth Backup
Cold backup of **high-resolution (1920×1088) monocular depth pseudo-labels** with their
source photographs — the high-res counterpart of the pipeline behind
[`androidli/photo-depth-pseudo-labels`](https://huggingface.co/datasets/androidli/photo-depth-pseudo-labels).
Published as-is so the exact arrays used in our experiments stay reproducible.
Depth labels were produced with a **Depth-Anything-V2** teacher over public photo-library
images (Wikimedia Commons / iNaturalist / Openverse / Pexels / Pixabay / Burst and
other collections).
## Layout
| Path | Content |
|---|---|
| `hires1920/cold/` | 148 shards, ~145 GB — current generation (`shard_s0_*.npz`) |
| `hires1920/oldpool/` | 61 shards, ~60 GB — older generation kept for lineage (`shard_old_s0_*.npz`) |
| `quota_test.bin` | 1 MB LFS upload-test artifact — ignore |
Each ~1 GB `.npz` shard is self-contained and holds ~512 samples as object arrays:
| Array | Content |
|---|---|
| `id` | sample identifier |
| `box` | bounding box / crop metadata |
| `rgb_jpg` | the source photograph, JPEG bytes |
| `gt_b2` | depth ground truth, **blosc2-compressed float16**, 1088×1920 |
```python
import io, blosc2, numpy as np
from PIL import Image
z = np.load("hires1920/cold/shard_s0_00030.npz", allow_pickle=True)
raw = z["rgb_jpg"][0] # JPEG bytes (object array)
rgb = np.asarray(Image.open(io.BytesIO(bytes(raw))).convert("RGB"))
gt = blosc2.decompress2(z["gt_b2"][0].tobytes())
depth = np.frombuffer(gt, dtype=np.float16).reshape(1088, 1920)
```
## Depth semantics
Depth-Anything raw convention: **larger value = farther, `0` = sky / invalid** (≈30% of
pixels can be 0). Relative depth only — no metric scale, no intrinsics.
## Licensing
No formal license is attached to this repository — it is published **as-is, for research
reproducibility only**. What the components are, so you can judge reuse yourself:
- **Depth arrays (`gt_b2`)** — our own model outputs, produced with a Depth-Anything-V2
teacher over the photos below.
- **Photographs (`rgb_jpg`)** — third-party images kept **as collected**, mixed licenses
per source, a large share collected without a recorded license. **Do not assume every
photo is redistributable.** Per-sample provenance for the overlapping pipeline is in
`photo-depth-pseudo-labels/metadata/` (`index.csv`, `attribution.csv`); check there
before re-sharing any bundled image.
If you want to reuse this data beyond research, or you are a rights holder and want
samples removed, **open an issue on this repo** and we'll sort it out promptly.
## Limitations
- `oldpool/` is a legacy generation: expect format/quality drift vs `cold/`.
- `shard_s0_*` naming is an internal convention, not a stable contract; treat shard
indices as opaque.
- No per-sample `license` column inside the shards — join against the photo-depth
metadata for licensing detail.