Energy-Aware NECO for Single-Pass Pixel-wise Out-of-Distribution Detection in Semantic Segmentation
Researchers have developed Energy-Aware NECO, a novel method for detecting out-of-distribution (OOD) data in semantic segmentation tasks, particularly for mobile robots. This single-pass approach combines a geometric ratio from decoder features with an Energy score, offering improved efficiency over methods like Monte Carlo Dropout. Evaluations on the miniMUAD dataset showed the hybrid score achieved an AUROC of 0.8539, surpassing existing baselines. AI
IMPACT Enhances reliability of AI systems in real-world, unpredictable environments by improving OOD detection.