AdapToPASS: Ambiguity-aware Adaptive Spherical Transformer for Panoramic Semantic Segmentation
Paper • 2608.29081 • Published
Datasets:
AdapToPASS-Stanford2d3d
A derived robustness dataset containing 1,413 source panoramas in each of 15 configurations: three severities (mild, robust, stress) and five families (rotation, scale, orientation, translation, viewpoint). These are synthetic transformations of the same scenes, not 21,195 independent observations. Images are 4096 × 2048.
There are 21,195 sample groups and 93,258 sample PNGs, plus one pair of shared root masks. The image package is approximately 190.28 GB. This release includes panoramic modalities only, not the source dataset's meshes, poses, normals, or perspective images. This is a derived release, not an official Stanford release.
License and access
The bundled original agreement governs the source data and derived information. LICENSE.txt is a convenience extraction; the original PDF controls. Use is limited to non-commercial academic purposes. Preserve the agreement and attribution when redistributing. Sections 2.4(b)–(c) specify recipient acceptance, acceptance records, and contact-information obligations for substantial redistribution.
The repository owner must enable gated access before making the data available. The form metadata above prepares the acceptance fields; it does not activate the Hub access gate. See PUBLISHING.md. No permissive license is granted to the data by this package.
Layout and manifests
README.md, LICENSE.pdf, LICENSE.txt, CITATION.bib
stanford2d3d_mask_common.png, stanford2d3d_mask_pretty.png
stanford2d3d_id2label.npy, stanford2d3d_colors.npy, name2label.json
<severity>/<family>/manifest.jsonl
<severity>/<family>/stanford2d3d/<area>/pano/<modality>/*.png
Each variant manifest contains 1,413 rows. All modality paths are POSIX paths relative to the dataset root, not the manifest directory. Fields include area, base_id, source_id, transform, and generation_seed. source_files records relative paths in the original Stanford dataset for provenance; these are not files to load from this release. Schema version is 2.
transform records rotation angles in degrees, dimensionless zoom, and, for translation and viewpoint families, effective translation in metres. generator_transform retains the original translation generator settings needed to reproduce the existing images; see the unit limitation below.
Modalities, labels and masks
rgb: colour image; convert to RGB when loading because the rotation, scale, and orientation outputs can retain source alpha channels.depth: uint16 radial depth PNG. The source encoding is 512 units per metre; divide by 512 for metres. Zero is padding/missing output, and 65535 is the source invalid sentinel. Rotation, scale, and orientation interpolate depth numerically and can blend invalid values into neighbours; masking does not fully correct these boundary artifacts.semantic: encoded instance IDs, not class colours. For RGB pixels with R=0, indexstanford2d3d_id2label.npyusingG*256+B. Invalid/red-channel-nonzero or out-of-range indices map to class 0. Apply the appropriate validity mask after decoding.semantic_pretty: visualization only; do not use it as class IDs.valid_mask: translation and viewpoint families only; 0 invalid, 255 valid after projection. Use this per-image mask for translation evaluation.
name2label.json maps 0 to unknown/ignore and 1–13 to beam, board, bookcase, ceiling, chair, clutter, column, door, floor, sofa, table, wall, window. stanford2d3d_colors.npy is a 14×3 uint8 class-colour table; stanford2d3d_id2label.npy has 9,817 uint8 entries. Both load with allow_pickle=False.
There is only one shared mask pair at the root. stanford2d3d_mask_common.png selects rows 309 through 1738 inclusive; the other rows are padded to zero. stanford2d3d_mask_pretty.png is the original source mask used to determine that crop band. The shared masks are not substitutes for translation valid_mask images or depth/semantic validity checks.
Transform definitions and limitations
Family names: scale applies spherical zoom; orientation combines rotation and spherical zoom; viewpoint combines rotation and depth-based translation. rotation and translation retain their original meanings. The numeric transform.zoom and zoom_phi_deg parameters retain their names and values.
The rotation, scale, and orientation pipeline crops the shared valid band, treats that band as the full ERP sphere, applies rotation and then optional spherical Möbius zoom, and pads back to full size. RGB/depth use bilinear interpolation; semantic images use nearest-neighbour. Zoom >1 denotes zoom-in. This cropped-band convention is part of this benchmark, not a physically exact rotation of the original full panorama.
Translation lifts valid radial depth onto rays, computes (points - translation) @ rotation.T, projects back to ERP, and splats to four neighbouring pixels with nearest-depth occlusion selection. It cannot synthesize unseen surfaces; holes are invalid and zero-filled. Seeds are 20260329 for rotation/scale/orientation and 20260410 for translation/viewpoint. Each source gets one sampled transform per configuration, not every combination of the parameter grid.
Historical translation depth-scale mismatch
The existing generator divided depth by 1000 although Stanford encodes 512 units per metre. Consequently the effective physical translations in these released images are 1000/512 = 1.953125 times the historical tx_m, ty_m, tz_m values. The release manifests correct transform to effective metres and preserve the historical values in generator_transform. Images have not been regenerated or rescaled. Interpret output depth in source units (512 per metre); do not describe the historical generator values as calibrated metre offsets. Float32 projection and rounding still reflect the historical computation.
Loading
Use ordinary file loading; the JSONL manifests are portable indexes and do not automatically configure a Hugging Face image viewer. After downloading the gated repository with your authenticated account:
from pathlib import Path
import json
import numpy as np
from PIL import Image
root = Path('/path/to/AdapToPASS-Stanford2d3d')
row = json.loads((root / 'mild/translation/manifest.jsonl').read_text().splitlines()[0])
rgb = np.array(Image.open(root / row['rgb']).convert('RGB'))
depth_raw = np.array(Image.open(root / row['depth']))
depth_metres = depth_raw.astype(np.float32) / 512.0
mask_path = root / row.get('valid_mask', 'stanford2d3d_mask_common.png')
valid = np.array(Image.open(mask_path).convert('L')) > 0
valid &= (depth_raw > 0) & (depth_raw != 65535)
sem = np.array(Image.open(root / row['semantic']).convert('RGB')).astype(np.int32)
id2label = np.load(root / 'stanford2d3d_id2label.npy', allow_pickle=False)
index = sem[..., 1] * 256 + sem[..., 2]
known = (sem[..., 0] == 0) & (index < len(id2label)) & valid
labels = np.zeros(index.shape, dtype=np.uint8)
labels[known] = id2label[index[known]]
Reproduction and validation
Per-sample generation settings are recorded in the manifests. The transformation utility is maintained separately; generator source and dependencies are not bundled in this release.
Completed data passed inventory checks across all variants and PNG signature, dimensions and end-marker checks across every PNG. Those checks are not a full CRC/pixel-decode scan of all images, nor a claim of scientific validation of the transform model. Release preparation separately tests loading representative images.
Source attribution
Source: Joint 2D-3D-Semantic Data for Indoor Scene Understanding, Iro Armeni, Sasha Sax, Amir R. Zamir, and Silvio Savarese (2017). Please cite CITATION.bib and describe these derived transformations and their limitations in any evaluation.
The official source documentation defines the depth encoding and semantic format. The software license in that upstream repository is distinct from the data agreement bundled here.
Citation
@article{debnath2026adaptopass,
title={AdapToPASS: Ambiguity-aware Adaptive Spherical Transformer for Panoramic Semantic Segmentation},
author={Debnath, Soumyaratna and Zhang, Weiming and Damodaran, Shriram and Xiao, Dingwen and Wang, Addison Lin},
journal={arXiv preprint arXiv:2608.29081},
year={2026}
}
- Downloads last month
- 154