nosmall/eval_our

Panorama Depth Benchmark

A unified evaluation suite for equirectangular (360°) depth estimation, assembled from the test splits of eight public panorama depth datasets. Every sample listed in <dataset>/test.txt has been copied into this repository — the loader never reads anything outside this folder.

• Modality: equirectangular RGB + metric depth (+ validity mask on Holo360D / ZuriPano)
• Total test samples: 15 766 across 12 subsets (2 outdoor-only, the rest indoor / nature / urban)
• Total size on disk: ≈ 61 GB
• License: each subset keeps its original license; see the per-dataset rows below for canonical sources.

---

Datasets

key	N	split	scene	RGB ext	depth ext	depth → meters	valid range (m)	mask file
stanford2d3d	373	test	indoor	png	png u16	v / 512.0	0.001 / 5.0	—
matterport3d	2 014	test	indoor	jpg	png u16	v / 4000.0	0.001 / 5.0	—
panosuncg	3 944	test	synthetic	png	png u8-RGB	v[...,0] * 10 / 255	0.001 / 5.0	—
structured3d	1 691	test	synthetic	png	png u16	v / 1000.0	0.001 / 5.0	—
metropolis	3 000	test	outdoor	jpg	png u16	v / 512.0	0.2 / 128.0	—
zuripano	100	test	outdoor	parquet	parquet	v_u16 * 200/65535	0.01 / 75.0	✓ (in parquet row)
holo360d_indoor	360	test	indoor	jpg	EXR	first channel is metric	0.01 / 10.0	✓ (mask/)
holo360d_outdoor	90	test	outdoor	jpg	EXR	first channel is metric	0.01 / 75.0	✓ (mask/)
replica360_2k	130	test	indoor	jpg	.dpt	Sintel float32 (metric)	0.001 / 10.0	—
panoinfinigen_indoor	4 392	test	indoor	png	.npy	raw float16 (metric)	0.001 / 10.0	—
panoinfinigen_nature	848	test	nature	png	.npy	raw float16 (metric)	0.001 / 100.0	—
panoinfinigen_urban	788	test	urban	png	.npy	raw float32 (metric)	0.001 / 500.0	—

Depth range semantics

valid range above is the validity mask gate: a depth pixel is valid iff min < d ≤ max and d is finite (AND-ed with the mask file where present). These values are aligned with the cvpr/benchmark evaluation harness's valid_from_depth so that both code paths select the same pixels.

Note that the cvpr harness additionally clamps both prediction and GT to a (usually identical) [min_depth, max_depth] window when computing metrics. For the seven datasets shared with that harness the two ranges coincide with one exception worth knowing about: Metropolis uses a validity mask of (0.2, 128] m (DA360's convention) but clamps metrics to [0.01, 100] m.

Datasets not present in the cvpr harness (replica360_2k, panoinfinigen_*) carry ranges chosen for the data itself, not copied from it.

For datasets without a mask file, validity is derived from the depth map alone (min_depth < d ≤ max_depth and finite). For Holo360D the mask/ JPEG also removes the photographer rig / tripod region (mask > 127 is valid). ZuriPano's mask is stored inline in the parquet row (mask field, L-mode PNG bytes).

Depth is stored as radial distance on all panorama datasets (not Z / planar depth) — do not pre-wrap pano frames with a perspective-depth model's output.

Image resolution

Native sizes vary per dataset. The loader does not resize by default; each row returns the native array. Common sizes:

• Stanford2D3D: 512 × 1024
• Matterport3D: 1024 × 2048
• PanoSUNCG: 1024 × 2048
• Structured3D: 512 × 1024
• Metropolis: 1024 × 2048
• Holo360D: 1024 × 2048 (RGB) / 1024 × 2048 (EXR)
• ZuriPano: 2048 × 4096
• Replica360-2k: 1024 × 2048
• PanoInfinigen (all configs): 1024 × 2048 (downsampled 2× from the 2048 × 4096 source; depth uses an exact [::2, ::2] stride so no interpolated values appear)

---

Directory layout

benchmark/
├── stanford2d3d/
│   ├── test.txt          # <rgb_rel> <depth_rel>
│   ├── rgb/<stem>.png
│   └── depth/<stem>.png
├── matterport3d/
│   └── ...
├── holo360d_indoor/
│   ├── test.txt          # <rgb_rel> <depth_rel> <mask_rel>
│   ├── rgb/<stem>.jpg
│   ├── depth/<stem>.exr
│   └── mask/<stem>.jpg
├── zuripano/
│   └── data/test-{00000..00004}-of-00005.parquet
├── _loader.py            # PyTorch Dataset: load_dataset_torch(root, name)
├── benchmark.py          # HuggingFace datasets BuilderConfig
├── _collect.py           # (dev) regenerate the tree from source dirs
└── README.md

Stems are flattened from the original scene / room / frame hierarchy with __ as a separator, e.g. area_5a__camera_<uuid>_hallway_1_frame_.... Each stem is globally unique within its dataset.

---

Loading

1. Plain Python / NumPy

from _loader import PanoDepthDataset
ds = PanoDepthDataset(root="/path/to/benchmark", name="stanford2d3d")
sample = ds[0]
# sample is a dict:
#   rgb:    uint8  [H, W, 3]
#   depth:  float32 [H, W]     metric distance in meters, radial
#   valid:  bool    [H, W]     after depth gating + optional mask
#   stem:   str                flattened id, unique within dataset

2. HuggingFace datasets

from datasets import load_dataset
ds = load_dataset("/path/to/benchmark", "stanford2d3d", split="test")

The HF script loads bytes lazily (Image(decode=False)); call .cast_column("rgb", Image()) etc. if you want decoded arrays.

---

Source attribution

• Stanford2D3D — Armeni et al., Joint 2D-3D-Semantic Data for Indoor Scene Understanding, 2017. (official 2D-3D-S release)
• Matterport3D — Chang et al., ICCV 2017. (panoramas provided by the PanDA benchmark re-release.)
• PanoSUNCG — Wang et al., SUNCG-based 360° renderings. Depth encoded into the R channel of an 8-bit PNG.
• Structured3D — Zheng et al., ECCV 2020. Only the panorama view and the raw-light rendering are used.
• Metropolis — DA360 (Depth-Anywhere) outdoor street-view set; see the DA360 project page.
• ZuriPano — PaGeR / prs-eth/ZuriPano on HuggingFace. 100 real outdoor LiDAR scans.
• Holo360D — a mixed real-world capture (12 indoor + 3 outdoor scenes); 360 / 90 frames linspace-sampled from each scene.
• Replica360-2k — 2K-resolution equirectangular renderings of the Replica indoor meshes (apartment / FRL / hotel / office / room). Depth stored as Sintel .dpt float32 (metric meters, no scale factor).
• PanoInfinigen — nyu-visionx/PanoInfinigen on HuggingFace, three scene configs (indoor / nature / urban) procedurally rendered with Infinigen. Ships train / val / test; only test is mirrored here. The original normals field is dropped (~75% of the source bytes and unused for depth evaluation), and the 4096 × 2048 frames are downsampled 2× to 2048 × 1024. Depth is a bare numpy array (not an image file) with metric meters directly — float16 for indoor/nature, float32 for urban.

Please cite each original paper when using the corresponding subset.

---

Regenerating

If a source dataset is updated or the split is extended, re-run from the repo root (one dataset at a time, or all):

python _collect.py                     # all
python _collect.py structured3d        # one

_collect.py is idempotent (skips files already copied with the same size) and writes a fresh test.txt each time.