Skelne: Personal Objects

Skelne helps blind and low-vision people distinguish the personal household objects that a general camera may treat as identical. A sighted helper teaches the phone several views of each item. Later, audio framing guides the camera and Skelne speaks the household name only when the match is clear.

TEACH THE DIFFERENCE THAT MATTERS

Generic labels can describe two items with the same broad word even when choosing the right one matters at home. Skelne learns the names your household actually uses. A helper can teach “Blue band box,” “Orange band box,” “Door keys,” or any other short, unambiguous label.

Guided capture checks light, motion, centering, object size, frame-edge contact, and likely camera obstruction before saving a reference. Several varied prototypes keep one angle from defining the whole item.

KEY FEATURES

• Personal object library with user-chosen spoken names
• Automatic capture of stable, centered reference views
• Fused color, edge, and visual-embedding comparison
• A distinct unknown cue when distance or separation evidence is weak
• Directional tones and optional spoken framing instructions
• Ordinary-scene recording to teach the model what not to name
• Corrections and targeted practice for repeatedly confused pairs
• Helper workbench that adapts across phones and large screens
• Recognition evidence with measured refusal runs
• Optional App Lock with Android biometrics or a dedicated App PIN
• One encrypted archive for moving the complete household library

NONVISUAL RECOGNITION

The recognition surface stays focused on one job. A large lower control starts or stops the camera. Tones guide direction, distance, light, and stillness. A name is spoken only after the same accepted result remains stable across consecutive frames.

If an object is far from every learned example, close to a runner-up, part of a confusion pair, or poorly framed, Skelne gives an unknown cue instead of filling the gap with a guess.

A HELPER WORKBENCH

The helper side separates teaching, object management, ordinary-scene evidence, corrections, transfer, and settings from the listening surface. Objects can receive targeted views or be removed with their photographs and recognition records. Evidence summaries show whether the current household set is separable and report measured false accepts.

LEARN FROM REAL CORRECTIONS

When a match is wrong, choose the object actually in view or mark that none was present. Repeated cross-object errors create a focused confusion prompt. Two more views of each object then expand the evidence and rebuild calibration.

MOVE THE HOUSEHOLD LIBRARY

Create one encrypted skelne archive containing learned names, photographs, embeddings, thresholds, corrections, and evidence. Android lets you choose its location. A separate 26-character code is required before another installation replaces its library with the verified archive. App PIN and biometric preferences are excluded.

ACCESSIBILITY AND CONTROL

Skelne uses large touch targets, high contrast, live-region updates, reduced motion, TalkBack-aware controls, and layouts that adapt to larger text. Spoken guidance can be turned off while directional framing tones remain available. The household library can be deleted at any time.

IMPORTANT LIMITS

Skelne reports only a household label taught by the user. It does not read text, describe a scene, verify product contents, or provide health or safety instructions. Lighting, camera quality, wear, reflections, obstruction, and insufficient teaching views can affect a match. Use the unknown result as a request for a better view, not as a failure to be bypassed.

SUPPORT

Questions, bug reports, and feature ideas can be sent from the in-app feedback action or to www.saifullah.ai@gmail.com.