TSONet

Official PyTorch implementation of Two-Stream Ordinal Network (TSONet) for monocular building height estimation from optical remote sensing imagery.

This repository is a cleaned release version of the project. Only the final TSONet pipeline and the ablation branches relevant to the paper are retained.

---

Overview

TSONet is a two-stream network for building height estimation. It jointly models:

• a footprint segmentation branch for building-aware representation learning, and
• a height estimation branch for dense building height prediction.

The two branches interact through a Cross-Stream Exchange Module (CSEM), while height estimation is performed with Feature-Enhanced Bin Refinement (FEBR).

Overview of the proposed TSONet framework.

---

Main Features

• Clean training and testing pipeline for TSONet
• Support for the final FEBR head
• Support for a DETR-style head as an ablation option
• Support for selected ablations:
  • removing CSEM
  • replacing FEBR with a DETR-style head in BinsFormer
  • removing the footprint branch
  • removing the bin head

---

Repository Structure

.
├── train.py                   # Training entry
├── test.py                    # Testing / inference entry
├── dataloader/
│   └── PHDataset.py           # Dataset loader
├── models/
│   ├── TSONet.py              # Main model definition
│   ├── encoder.py             # Encoder backbone
│   ├── decoder.py             # Two-stream decoder and CSEM
│   ├── bins_head.py           # FEBR / DETR height heads
│   └── __init__.py            # Model factory
├── options/
│   ├── base_options.py
│   ├── train_options.py
│   └── test_options.py
└── utils/
    ├── losses.py              # Loss functions used by the released TSONet
    ├── metrics.py             # Evaluation metrics
    └── writers.py             # Prediction saving utilities

---

Dataset Preparation

The related dataset, PHDataset, is available at:

https://huggingface.co/datasets/Yanjiao-WHU/PHDataset

The dataset is organized as follows:

data_root/
├── train/
├── val/
├── test/
└── splits/
    ├── train.txt
    ├── val.txt
    └── test.txt

---

Citation

If you find this repository useful, please cite the corresponding paper, which is currently under review and available in arXiv:

@misc{song2026monocularbuildingheightestimation,
      title={Monocular Building Height Estimation from PhiSat-2 Imagery: Dataset and Method}, 
      author={Yanjiao Song and Bowen Cai and Timo Balz and Zhenfeng Shao and Neema Simon Sumari and James Magidi and Walter Musakwa},
      year={2026},
      eprint={2603.29245},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.29245}, 
}