Gaussian_YOLOv3

Why we included this project

This is the reference implementation of the ICCV 2019 Gaussian YOLOv3 paper, which extends YOLOv3 by modeling bounding-box localization uncertainty as a Gaussian. Instead of just predicting a box, the detector also reports how confident it is in that box's position, which is useful in safety-sensitive settings like autonomous driving. The code builds on the official Darknet YOLOv3 codebase, so anyone who has trained YOLO before can adapt it to their own data without much friction. It also ships a pretrained weight file validated on the Berkeley DeepDrive dataset, giving a working starting point rather than a paper you have to reimplement from scratch. For teams comparing real-time detectors that have to balance speed against reliability, this is a good reference for adding uncertainty estimates to an existing detector.