RampNet v1.0 — ICCV'25 CV4A11y (paper state)

This release preserves the exact code state (commit 8783157) used to produce the results in our paper, for scientific replicability. No code changes have been made since publication; all future development happens after this tag.

RampNet: A Two-Stage Pipeline for Bootstrapping Curb Ramp Detection in Streetscape Images from Open Government Metadata
John S. O'Meara, Jared Hwang, Zeyu Wang, Michael Saugstad, Jon E. Froehlich
ICCV'25 Workshop on Vision Foundation Models and Generative AI for Accessibility (CV4A11y)
📄 Paper: https://arxiv.org/abs/2508.09415

Published artifacts

• 🗃️ Dataset (214,376 panoramas / 849,895 labels): https://huggingface.co/datasets/projectsidewalk/rampnet-dataset
• 🤖 Trained Stage-2 model (AutoModel, trust_remote_code): https://huggingface.co/projectsidewalk/rampnet-model
• 🛠️ Colab demo: linked from the README

Known limitations of this snapshot (to be addressed post-release; see open issues)

• stage_two/demo.py and stage_two/evaluate.py reference a local checkpoint checkpoints/epoch_1_step_9378.pth that was never distributed (#1). The published Hugging Face model contains the trained weights — use the AutoModel example in the README for inference.
• environment.yml is a full linux-64 conda export and will not solve on other platforms; the README's "install CUDA 11.8" note does not match the pinned cu126 builds (#6).
• Detection thresholds differ across the README example (0.5), demo.py (0.4), and evaluate.py (0.0 — intentional: it sweeps all confidences to produce PR curves). Guidance on choosing an operating point is forthcoming (#7); as a starting point, 0.55 gives P=0.938 / R=0.935 on the 1k-panorama manual gold set with horizontal-flip TTA (see stage_two/evaluation_results/pr_rc_vs_c_data_manual_r0.022_pt0.0.csv).

Citation

@inproceedings{omeara2025rampnet,
  author    = {John S. O'Meara and Jared Hwang and Zeyu Wang and Michael Saugstad and Jon E. Froehlich},
  title     = {{RampNet: A Two-Stage Pipeline for Bootstrapping Curb Ramp Detection in Streetscape Images from Open Government Metadata}},
  booktitle = {{ICCV'25 Workshop on Vision Foundation Models and Generative AI for Accessibility: Challenges and Opportunities (ICCV 2025 Workshop)}},
  year      = {2025},
  doi       = {https://doi.org/10.48550/arXiv.2508.09415},
  url       = {https://cv4a11y.github.io/ICCV2025/index.html},
  note      = {DOI: forthcoming}
}