Lens — Sensor-Aware Image Acquisition under Shift

Stabilized recognition accuracy under natural distribution shift via adaptive sensor control — validated across multiple backbones.

Outcome. Prototyped a sensor-control loop that tunes capture parameters to scene/domain shift, yielding more stable accuracy on natural shift sets.

Role: Research member · Dates: Mar–Dec 2024 · Stack: Python, PyTorch, OpenCV, ImageMagick
Context: AIoT Group @ SNU

Highlights

  • Designed offline→online eval parity on real natural-shift datasets; tracked QoQ regressions with regression-safe comparisons.
  • Authored a data collection protocol to expose failure modes across illumination/ISO/exposure; automated labeling + metadata.
  • Demonstrated accuracy stability gains vs. fixed capture baselines across multiple backbones (CLIP-like and CNN).
giscus comments misconfigured

Please follow instructions at http://giscus.app and update your giscus configuration.