sijiey/City-Landscape-In-Sight

City Landscape In Sight — Window View Perception (Images & Models)

This dataset hosts the large binary artefacts for the paper "City landscape in sight: A crowdsourced framework for unlocking urban-scale window view perceptions from real estate imagery." It is the companion of the code repository on GitHub:

• Paper (arXiv): https://arxiv.org/abs/2606.15198
• Code & derived data (GitHub): https://github.com/Sijie-Yang/City-Landscape-In-Sight
• Images & trained weights (this dataset): https://huggingface.co/datasets/sijiey/City-Landscape-In-Sight

The framework uses 12,334 window view images (WVIs) collected from living-room-balcony viewpoints in real-estate listings in Wuhan, China, with perception labels crowdsourced through a non-immersive virtual-reality survey (27,477 pairwise comparisons from 304 participants on 499 sampled WVIs across six perceptual dimensions).

Contents

data/
├── training_images/
│   ├── WVI_Processed/        # 499 processed (256 px, masked + cropped) surveyed WVIs
│   └── WVI_Segmentation/     # semantic segmentation masks for the surveyed WVIs
└── inference_images/
    └── WVI_Processed.zip     # 12,334 processed citywide WVIs (zipped; one dir exceeds the 10k-file limit)
model_outputs/
├── kfold_spatial/           # trained weights (spatial-block CV) — best_model_{dim}.pth + aux_scaler_{dim}.joblib
└── kfold_random/            # trained weights (random stratified CV) — best_model_{dim}.pth + aux_scaler_{dim}.joblib

Perception dimensions: prefer, monotonous, quiet, extensive, vivid, oppressive.

Trained weights. Each best_model_{dim}.pth (~94 MB) is a ResNet-50 backbone with a multi-pathway tabular head, trained per dimension. For the citywide perception maps, a single retained model per dimension — the best-performing spatial-block fold checkpoint — is applied to all 12,334 WVIs (the 5-fold spatial CV is used only to estimate out-of-sample accuracy). Load with the matching aux_scaler_{dim}.joblib for the tabular features; see code_4_wvi_perception_model_train_predict.ipynb in the GitHub repository.

Not included

The raw window view images (WVI_Original) obtained from the real-estate platform are not redistributed owing to platform licensing restrictions. The image pre-processing scripts in the GitHub repository (Code 2) regenerate the processed imagery from source.

Usage

The citywide images are provided as a single WVI_Processed.zip; extract it to data/inference_images/WVI_Processed/ to obtain the 12,334 JPEGs expected by the code.

from huggingface_hub import snapshot_download

# Download everything
local_dir = snapshot_download(
    repo_id="sijiey/City-Landscape-In-Sight",
    repo_type="dataset",
)

# Or a single trained model
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download(
    repo_id="sijiey/City-Landscape-In-Sight",
    repo_type="dataset",
    filename="model_outputs/kfold_spatial/best_model_prefer.pth",
)

License

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

Peng, C., Yang, S., Liu, A., Xiang, Y., Zhou, Z., & Biljecki, F. (2026). City landscape in sight: A crowdsourced framework for unlocking urban-scale window view perceptions from real estate imagery. arXiv:2606.15198. https://arxiv.org/abs/2606.15198

@article{peng2026citylandscape,
  title        = {City landscape in sight: A crowdsourced framework for unlocking urban-scale window view perceptions from real estate imagery},
  author       = {Peng, Chucai and Yang, Sijie and Liu, Ang and Xiang, Yang and Zhou, Zhixiang and Biljecki, Filip},
  journal      = {arXiv preprint arXiv:2606.15198},
  year         = {2026},
  doi          = {10.48550/arXiv.2606.15198},
  url          = {https://arxiv.org/abs/2606.15198}
}