Existing vision datasets largely overlook the small, accessibility-critical objects that blind and low-vision (BLV) users rely on when navigating, such as audible pedestrian signals, door-activation buttons, and handrails. The few datasets that do include them tend to have limited class coverage, inconsistent annotations, and poor diversity.

To fill that gap, the authors introduce NavAble, a dataset spanning 11 accessibility-critical object classes. It combines 41K curated open-source images with 8K newly collected, densely annotated real-world images. They also built a synthetic data generation pipeline that renders high-fidelity 3D assets across 37 environments, producing 452K frames with rich ground-truth annotations from 565 assets. The pipeline automates web image crawling, VLM-based filtering, SAM 3D-based asset generation, and camera trajectory configuration, reducing manual effort.

Experiments show that fine-tuning existing segmentation models on NavAble yields reliable accessibility-object segmentation, exceeding 80% mIoU on the held-out test set. Adding synthetic data improves performance by up to +3.5 mIoU over real-world-only fine-tuning. Validation on egocentric data from two mobility-assistive robots also shows improved perception reliability in real-world navigation scenarios, suggesting the dataset and pipeline offer a scalable foundation for BLV navigation perception.