HIndoor-8K

HIndoor-8K is the first metrically calibrated real-world benchmark of indoor RGB–D panoramas at native 8192×4096 (8K) resolution. It provides 49 equirectangular RGB panoramas, each paired with a sparse metric depth map rendered from a real LiDAR point cloud, across 5 representative indoor environments.

Released as a community resource for high-resolution 360° depth estimation.

---

Contents

HIndoor-8K/
├── README.md
├── ich/          # corridor
│   ├── rgb/      <timestamp>.jpg     RGB equirectangular panorama (8192×4096)
│   └── depth/    <timestamp>.png     sparse metric depth (16-bit, millimetres)
├── interrato/    # basement
├── pt/           # project room
├── sala/         # hall / meeting room
└── unibs/        # laboratory

For every rgb/<timestamp>.jpg there is a matching depth/<timestamp>.png with the same filename stem (the capture timestamp).

Scene	Environment	# pairs
ich	corridor	7
interrato	basement	8
pt	project room	10
sala	hall / meeting room	14
unibs	laboratory	10
Total		49

---

Data format

RGB — rgb/<timestamp>.jpg

• Real spherical photograph, equirectangular projection, 8192×4096, 3-channel sRGB.
• Real illumination, HDR/exposure variation, and sensor noise are preserved (not retouched).

Depth — depth/<timestamp>.png

• 16-bit single-channel PNG (I;16, unsigned), 8192×4096, pixel-aligned to the RGB.
• Stored in millimetres: depth_metres = png_value / 1000.0.
• Sparse. The map is a projection of a real LiDAR point cloud, so a large fraction of pixels have no valid return (specular, distant, and grazing-angle surfaces, plus the LiDAR's intrinsic angular sparsity). Invalid pixels are stored as 0. Typical valid coverage is ≈ 15–20 %.
• No densification / inpainting is applied — this is the raw rendered depth.

Reading depth (Python)

import numpy as np
from PIL import Image
Image.MAX_IMAGE_PIXELS = None

d_mm  = np.asarray(Image.open("depth/<timestamp>.png"), dtype=np.float32)  # millimetres
valid = d_mm > 0                       # validity mask (0 = no LiDAR return)
d_m   = d_mm / 1000.0                  # metres

Evaluation convention

Compute all metrics only over valid pixels (depth > 0). In our paper we additionally cap the maximum depth at 16 m for the reported HIndoor-8K numbers; adopt the same cap to reproduce them:

mask = (d_m > 0) & np.isfinite(d_m) & (d_m <= 16.0)

Because the depth is relative-free / metric but sparse, predictions from relative-depth methods should be aligned to the ground truth (e.g. per-image median-ratio scale alignment) before computing error metrics.

---

Acquisition

Data is captured with a professional indoor mobile-mapping platform carrying rigidly coupled, mutually calibrated sensors:

• LiDAR: dual Hesai XT32-class scanners (10–20 Hz), 32 channels, 31–40° vertical FoV, up to 0.1–0.2° angular resolution, 120 m (MT1) – 300 m (M2X) range. Timestamped 3D points with reflectance / return-intensity.
• 360° RGB camera: a spherical camera rigidly mounted to the LiDAR rig, capturing equirectangular panoramas at 8192×4096 (photo mode), with fixed intrinsics and a known optical centre relative to the LiDAR.
• IMU: 3-axis accelerometer + gyroscope for SLAM-based trajectory refinement.

All sensors share a common reference frame. The pipeline outputs (i) a refined ~10 Hz trajectory, (ii) synchronized LiDAR scans, and (iii) timestamped RGB panoramas.

Depth rendering

Each RGB panorama's depth is produced by off-screen spherical rendering of the fused LiDAR point cloud, so the RGB stays a real photograph while depth comes from real geometry:

1. Pose interpolation — the 10 Hz trajectory is interpolated to the exact RGB timestamp.
2. Spherical projection — each 3D LiDAR point is projected to the spherical camera (azimuth / elevation) and mapped to equirectangular coordinates.
3. Depth assignment — each pixel stores the distance to the closest point along its ray; occluded or inconsistent samples are discarded. Pixels that receive no point remain 0 (→ the map is sparse).
4. Encoding — depth is written as a 16-bit millimetre PNG at native resolution.

---

Optional densification

If a dense map is required, the sparse depth can be completed with an off-the-shelf depth-completion model (e.g. LDCM, Large Depth Completion Model from Sparse Observations, ICLR 2026). We deliberately ship the raw sparse depth so that users may choose their own completion (or none). All benchmark numbers in the paper are computed on the raw sparse depth with a validity mask.

---

Citation

If you use HIndoor-8K, please cite the FRED paper (bibentry to be added on release):

@article{fred_hindoor8k,
  title  = {FRED: Full-Resolution Equirectangular Depth Estimation, and the HIndoor-8K Benchmark},
  author = {Shah, U. and Pintore, G. and Tukur, M. and Zahoor, A. and Schneider, J. and
            Sgrenzaroli, M. and Vassena, G. and V\'azquez, P.P. and Gobbetti, E. and Agus, M.},
  journal= {Graphical Models},
  year   = {2026}
}

License