nosmall/pano_train

Panorama Depth — Structured3D & PanoInfinigen

Two equirectangular (360°) RGB–depth collections, repacked for distribution.

• Datasets: 2 — Structured3D, PanoInfinigen
• Samples: 36 938 RGB–depth pairs (21 765 Structured3D + 15 173 PanoInfinigen)
• Total size: ≈ 60 GB
• Modality: equirectangular RGB + metric depth
• Camera files: none are included — and none are needed (see below)
• Licence: each collection keeps its original licence and attribution

---

Layout

Structured3D/
  Structure3D_train.txt        # 18 298 pairs
  Structure3D_val.txt          #  1 776 pairs
  Structure3D_test.txt         #  1 691 pairs
  scene_00000/2D_rendering/485142/panorama/full/
      rgb_rawlight.png
      depth.png

PanoInfinigen/
  data/
    urban/train-NNNNN.parquet    # 230 shards, 24 rows each
    nature/train-NNNNN.parquet   # 431 shards, 24 rows each

Split lines are <rgb_rel> <depth_rel>, both relative to Structured3D/.

---

Structured3D

scenes	3 500
scene split	00000–02999 / 03000–03249 / 03250–03499
files per pair	rgb_rawlight.png + depth.png
size	≈ 24 GB

The upstream release renders eight images per panoramic view (albedo, depth, instance, normal, rgb_coldlight, rgb_rawlight, rgb_warmlight, semantic). This copy keeps the RGB panorama and its depth map only; the remaining six renders account for ~95% of the original 496 GB.

depth.png is uint16, millimetres: depth_m = v / 1000.0.

Ancillary annotation files (annotation_3d.json, bbox_3d.json and the annotations/ archives) are not included.

---

PanoInfinigen

Source: prs-eth/PanoInfinigen (urban + nature subsets).

Upstream shards carry ~89 MB per row across three fields. This copy keeps the two image/geometry fields and drops one:

field	upstream	here
image	PNG 4096×2048 RGB, ~9 MB	JPEG q95 2048×1024
depth	.npy float16 2048×4096, 16–32 MB	.npy float32 1024×2048
normals	.npy float16 2048×4096×3, ~48 MB	not included

normals is 54% of the upstream payload. Spatial downsampling by 2× uses an exact [::2, ::2] stride, so no interpolated values are introduced.

Net effect: 643 GB → 37 GB.

Parquet row format

The schema matches upstream, so code written against the original file works unchanged:

image    struct<bytes: binary, path: string>   # JPEG bytes; `path` is null
depth    binary                                # plain .npy bytes, float32, (1024, 2048), metres

import io, pandas as pd, numpy as np
from PIL import Image

df = pd.read_parquet("PanoInfinigen/data/urban/train-00000.parquet")
row = df.iloc[0]

rgb   = np.asarray(Image.open(io.BytesIO(row["image"]["bytes"])).convert("RGB"), np.uint8)  # (1024, 2048, 3)
depth = np.load(io.BytesIO(row["depth"])).astype(np.float32)                                # (1024, 2048), m

depth is stored as uncompressed .npy so np.load(io.BytesIO(...)) works directly. Compression is applied at the parquet level (zstd), which takes each row from 8.6 MB to 3.8 MB.

Depth range

Measured range of the depth maps in the urban and nature subsets:

subset	min	max
urban	0.81 m	500 m
nature	0.30 m	75 m

The 500 m value in urban is a sky sentinel: pixels at that distance mark the sky, not a measurable surface. Exclude it by clamping below 500 m.

Rows are 24 per shard, matching upstream granularity.

---

Depth convention

Every sample is an equirectangular panorama. Depth is stored as radial distance from the camera centre, not planar / Z depth. ERP pixel coordinates map directly to a spherical direction (θ, φ), so the collections carry no intrinsics or extrinsics — there is no perspective projection to invert, and no per-scene camera parameters exist in these files.

---

Per-sample normalisation

Not applied here — both collections ship raw metric depth. The values below are the per-image convention used when reading them:

per-image log-depth mapped to [-1, 1] using the 2/98 percentile window of the valid pixels.

---

Attribution

Please cite the original works:

• Structured3D — Jia et al., Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling (ECCV 2020).
• PanoInfinigen — Chen et al., PanoInfinigen (prs-eth/PanoInfinigen).

Each collection remains under its original licence; see the respective sources for terms.