ZFTurbo/Weighted-Boxes-Fusion

What it solves

This project provides a way to combine (ensemble) bounding box predictions from multiple object detection models into a single, more accurate set of predictions. It addresses the problem of redundant or overlapping boxes when multiple models are used together to improve detection performance.

How it works

The library implements several algorithms for merging boxes, including Non-maximum Suppression (NMS), Soft-NMS, Non-maximum weighted (NMW), and the primary Weighted Boxes Fusion (WBF) method. WBF improves upon traditional methods by fusing boxes based on their confidence scores and weights assigned to each model, rather than simply discarding overlapping boxes.

Who it’s for

It is designed for developers and researchers working with object detection models in computer vision, as well as those performing Named-entity recognition (NER) in NLP tasks using the 1D version of the tool.

Highlights

• Supports 1D, 2D, and 3D bounding box fusion.
• Includes multiple ensembling methods: NMS, Soft-NMS, NMW, and WBF.
• Compatible with Python 3, Numpy, and Numba for performance.
• Provides benchmarks for Open Images, COCO, and NLP datasets.