Simulated Prosthetic Vision
Master's thesis playground: a PC-based Python pipeline that takes an RGB-D source, detects walkable space, encodes the scene, and renders it as a simulated phosphene percept with latency metrics — the platform for a human study comparing a raw depth encoder against a simplified, hazard-aware one.
focus
Designing what a person with a retinal implant would actually perceive, then measuring whether a smarter encoding (suppress the floor, light up obstacles by proximity, pulse drop-offs, mark the deepest walkable direction) beats raw depth under an identical simulated implant.
milestones
01
RGB-D → RANSAC floor + height rules → encoder → phosphene renderer (jitter, dropout, brightness levels, afterglow)
02
Two experimental encoders sharing one simulated implant so percept differences come from encoding alone
03
TinySegNet: ~70K-parameter depthwise-separable U-Net predicting walkability from inverse depth + validity, trained on procedurally generated corridors
04
Synthetic data generator, training/evaluation CLI (IoU + latency vs. rule baseline), NYU Depth V2 support
05
Live viewer with debug/percept toggles, headless capture mode, and a pytest suite