KF-LV 1.0: a low-vision test set for perception models
With Frothly’s Blackwall programme. 27,500 annotated scenes recorded at eye height in poor light, on stairs, at crossings and in crowds, for testing whether the perception software in assistive devices fails in the places that matter most to blind and partially sighted people.
KF-LV is a set of 27,500 scenes recorded at eye height in the places where perception software in assistive devices most often fails blind and partially sighted people: on stairs, at crossings, in crowds, at night and in rain. Each scene is annotated by hand, including every white cane, guide dog and harness.
The scenes were recorded by sighted volunteers and by volunteers with low vision, in Chiba, Osaka, Nairobi and Lalitpur, with the consent of everyone identifiable in them. Faces have been blurred.
Key findings
- Released under a data use agreement that forbids any attempt to identify anyone in the scenes.
- Used by Blackwall in its evaluation of SHIKAI 4, published in November 2025.
- A second release, with more scenes from rural roads, is planned for 2027.
How to cite
Chiba Centre for Visual Science, with Frothly’s Blackwall programme (2025). KF-LV 1.0: a low-vision test set for perception models. KF-D-2025-01. Tokyo: Kiroshi Foundation. kiroshi.org/research/publications/kf-d-2025-01/
For the full text, write to research@kiroshi.org.