PAVE: Pedestrian Accessibility and Visual-grounded Evaluation

PAVE is a structured vision–language dataset introduced in the CVPR 2026 paper:

WalkGPT: Grounded Vision–Language Conversation with Depth-Aware Segmentation for Pedestrian Navigation
(Accepted at CVPR 2026)
Paper link: https://arxiv.org/abs/2603.10703

PAVE is a spatially grounded VQA benchmark for accessibility-aware reasoning in real-world pedestrian environments, unifying language understanding, pixel-level grounding, and depth-aware navigation guidance.

---

Overview

PAVE provides structured annotations built on top of the publicly available SANPO dataset from Google Research:

https://github.com/google-research-datasets/sanpo_dataset

We sincerely thank the SANPO authors for making their dataset publicly available.
PAVE does not redistribute SANPO images. Instead, it provides:

• Question–Answer pairs
• Accessibility reasoning annotations
• Segmentation references
• Distance annotations
• Structured grounding tags

Users must independently download SANPO under its original license to use PAVE.

---

Dataset Structure

The repository contains:

• PAVE.jsonl — Full dataset
• PAVE_train.jsonl — Official train split
• PAVE_val.jsonl — Official validation split
• PAVE_train85.jsonl — Session-balanced training subset (85 sessions)
• PAVE_val85.jsonl — Session-balanced evaluation subset (6 sessions)
• PAVE_labelmap.json — Ontology and accessibility label definitions

---

Official Splits

To reduce redundancy while preserving scene-level diversity, we uniformly sample 100 frames per session.

• 85 sessions used for training (~8.5k frames)
• 6 sessions held out for evaluation (~600 frames)

These splits correspond to the setup described in the WalkGPT paper.

For full training, use:

• PAVE_train.jsonl
• PAVE_val.jsonl

For session-balanced experiments (as in the paper):

• PAVE_train85.jsonl
• PAVE_val85.jsonl

---

Annotation Format

Each JSON entry contains:

• Reference to SANPO image file
• Question
• Structured reasoning output
• Segmentation grounding tags [p] ... [/p][SEG]
• Distance annotations [distance] ... [/distance]
• Accessibility assessment tags [assessment] ... [/assessment]

The dataset is designed to evaluate:

• Grounded multimodal reasoning
• Accessibility-aware scene understanding
• Depth-aware spatial interpretation
• Structured segmentation-language alignment

---

Accessibility Ontology

The file PAVE_labelmap.json defines:

• Semantic classes
• Mapping of classes to accessibility categories
• Definitions of accessible vs non-accessible elements

Accessibility categories are defined from a pedestrian navigation perspective as described in the WalkGPT paper.

For detailed ontology design and evaluation protocol, please refer to the paper.

---

Relationship to SANPO

PAVE annotations are derived from the SANPO dataset.

• Images, depth maps, and base masks are not redistributed here.
• Users must obtain SANPO independently.
• PAVE provides annotation layers on top of SANPO.

Please cite SANPO when using the underlying images.

---

Intended Use

PAVE is intended for:

• Research in grounded vision–language models
• Accessibility-aware navigation systems
• Spatial reasoning benchmarks
• Multimodal segmentation research

---

For a detailed discussion of limitations and ethical considerations, see the WalkGPT paper.

---

License

PAVE annotations are released under:

Creative Commons Attribution 4.0 (CC BY 4.0)

This license applies to the annotation files only.
SANPO images are governed by their original license.

---

Citation

If you use PAVE, please cite:

@misc{sultan2026walkgptgroundedvisionlanguageconversation,
      title={WalkGPT: Grounded Vision-Language Conversation with Depth-Aware Segmentation for Pedestrian Navigation}, 
      author={Rafi Ibn Sultan and Hui Zhu and Xiangyu Zhou and Chengyin Li and Prashant Khanduri and Marco Brocanelli and Dongxiao Zhu},
      year={2026},
      eprint={2603.10703},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.10703}, 
}