fizyr/keras-retinanet
Keras Retinanet
This project is a deep learning implementation of the RetinaNet architecture for detecting and classifying objects within images. Built as a Keras object detection framework and a TensorFlow computer vision tool, it provides a complete neural network implementation based on the RetinaNet paper.
The framework includes specialized components such as a Feature Pyramid Network and a focal loss function to handle object detection. It features a configurable backbone architecture and anchor-based bounding boxes to predict object locations across varying scales and aspect ratios.
The toolset covers the end-to-end workflow for computer vision, including training routines, performance evaluation, and model inference deployment. It provides data management utilities for importing and debugging image annotations from CSV and Pascal VOC formats, as well as tools for converting trained models into different formats for deployment.
Features
- Object Detection - Implements a complete system for identifying and locating objects within images using bounding boxes and classification.
- Object Detection Training - Implements a full RetinaNet training pipeline to teach neural networks to localize and classify objects in images.
- Anchor Box Detections - Implements anchor box mechanisms and non-max suppression to predict precise object locations.
- Multi-Scale Feature Pyramids - Uses a Feature Pyramid Network to aggregate hierarchical representations for detecting objects of varying sizes.
- Focal Loss Calculators - Implements focal loss to down-weight easy examples and focus training on difficult object observations.
- Neural Network Model Implementations - Provides a practical implementation of the RetinaNet architecture, including a Feature Pyramid Network and focal loss.
- TensorFlow Model Development - Provides a comprehensive toolkit for developing and training object detection models within the TensorFlow ecosystem.
- Object Detection Frameworks - Provides a comprehensive toolkit for implementing the RetinaNet object detection architecture.
- Anchor Optimizers - Includes utilities to optimize anchor box dimensions based on dataset characteristics for improved detection accuracy.
- Image Annotation Integration - Integrates user-provided image annotations from CSV and Pascal VOC into the training pipeline.
- Inference Model Deployment - Converts trained networks into optimized formats for hardware execution and object detection inference.
- Model Conversion Tools - Transforms trained regression and classification networks into optimized formats suitable for object detection inference.
- Model Performance Evaluators - Quantifies the accuracy and precision of detection models by comparing predictions against ground truth test datasets.
- Pre- and Post-Inference Transformations - Ships post-inference transformations that decode raw tensor outputs into human-readable bounding boxes.
- Modular Backbone Architectures - Provides a modular backbone architecture allowing the base feature extraction network to be swapped.
- Keras Model Implementations - Implements the detection architecture specifically using the Keras API for a practical object detection framework.
- Vision Model Evaluation - Provides frameworks for assessing the accuracy and performance of trained object detection models.
- Model Export Formats - Transforms trained models into different formats to ensure compatibility with various deployment runtimes.
