Explainable Road Hazard Detection via Pixel-wise Uncertainty Analysis

Dec 11, 2024 · 1 min read


Pixel‑wise logit variance drives a segmentation network that spots debris and potholes while quantifying risk for autonomous vehicles.

Highlights

  • Pixel-wise logit variance produces interpretable uncertainty maps
  • Iterative background highlighting refines anomaly localization
  • Evaluated for road-anomaly segmentation in the published ICRA study

Publications

  • Road anomaly segmentation based on pixel‑wise logit variance with iterative background highlighting, ICRA 2023
PhD Thesis XAI Anomaly Detection Computer Vision
Dongkun Lee
Authors
Dongkun Lee
Senior Staff Engineer | LLM Inference Optimization | Model Compression

I am a Senior Staff Engineer working on efficient AI systems, with a focus on LLM inference optimization and model compression.

My work spans quantization, pruning, knowledge distillation, sparsity, and decoding algorithms, with an emphasis on reducing deployment cost while preserving model quality.

I hold a Ph.D. in Computer Science from KAIST, with a research background in out-of-distribution detection, anomaly segmentation, uncertainty estimation, generative modeling, and computer vision.