Likelihood-ratio estimation for unknown objects
An unknown estimation module for detecting out-of-distribution regions using foundation models and likelihood ratios.
Overview
The unknown estimation module (UEM) detects out-of-distribution objects using likelihood-ratio scores on top of a semantic segmentation model. The implementation includes foundation-model segmentation heads, model checkpoints, and separate training stages for the segmentor and unknown estimation module.
The repository provides configurations and evaluation scripts for road-anomaly and Fishyscapes benchmarks.
Resources
Technologies
Python · PyTorch · Semantic segmentation · Uncertainty estimation