ZFTurbo/Weighted-Boxes-Fusion

Set of methods to ensemble boxes from different object detection models, including implementation of "Weighted boxes fusion (WBF)" method.

What it solves

It addresses the problem of ensembling predictions from multiple object detection models. When different models predict multiple overlapping bounding boxes for the same object, this library provides methods to merge those boxes into a single, more accurate prediction.

How it works

The library implements several box-fusion algorithms that take normalized coordinates, confidence scores, and labels from multiple models as input. It supports:

  • Weighted Boxes Fusion (WBF): A method that fuses boxes based on their confidence scores and weights assigned to each model.
  • Non-maximum Suppression (NMS): A standard method for removing redundant overlapping boxes.
  • Soft-NMS: A variation of NMS that decays the confidence of overlapping boxes rather than deleting them.
  • Non-maximum Weighted (NMW): An alternative weighted approach to box merging.

Who it’s for

It is designed for computer vision engineers and data scientists who use object detection models and want to improve their final prediction accuracy by combining the outputs of multiple models.

Highlights

  • Supports 2D, 3D, 1D (useful for NLP Named Entity Recognition), rotated, and quadrangle bounding boxes.
  • Includes implementations of NMS, Soft-NMS, NMW, and WBF.
  • Provides benchmarks for Open Images, COCO, and NLP datasets.
  • Easy installation via pip install ensemble-boxes.

Related

  • Project
    yzhao062/combo

    combo is a Python toolbox for combining machine learning models and scores. It offers unified APIs for ensemble techniques like stacking, dynamic classifier/ensemble selection, and outlier detector fusion, supporting classification, clustering, and anomaly detection with models from scikit-learn, XGBoost, LightGBM, and pyod.

  • Project
    Koldim2001/YOLO-Patch-Based-Inference

    A Python library that enables patch-based inference for YOLO models to improve the detection of small objects in large images through image tiling and result consolidation.

  • Project
    huggingface/evaluate

    A library for standardized evaluation and comparison of machine learning models, providing a wide range of metrics across NLP and Computer Vision.

  • Project
    Mayitzin/ahrs

    A pure Python library for estimating the orientation of mobile systems using various attitude and heading reference system (AHRS) algorithms.

  • Project
    rasbt/mlxtend

    A Python library providing machine learning extensions for data science tasks, including ensemble methods, feature selection, and visualization utilities.