This repository provides an implementation of EfficientDet, a family of scalable and efficient object detection models, for use with Keras and TensorFlow. It targets researchers and practitioners in computer vision who need a flexible and performant object detection solution. The project offers pre-trained weights and synchronization with the official Google implementation.

How It Works

This implementation leverages the EfficientNet backbone for feature extraction and incorporates the BiFPN (Bi-directional Feature Pyramid Network) for multi-scale feature fusion, a core innovation of EfficientDet. This combination allows for improved accuracy and efficiency across a range of model scales (phi 0-6), enabling users to balance performance and computational cost.

Quick Start & Requirements

• Install via pip.
• Requires Python 3.x, TensorFlow, and Keras.
• Pre-trained EfficientNet weights (ImageNet) and EfficientDet weights (COCO) are available for download.
• Training and evaluation scripts are provided for Pascal VOC and MSCOCO datasets.
• Official documentation and demo links are not explicitly provided in the README.

Highlighted Details

• Synchronized with the official Google automl implementation.
• Supports anchor-free detection (SAPD) for improved speed and smaller model size.
• Includes support for quadrangle detection.
• Provides benchmark results for Pascal VOC (mAP 50 up to 0.8029 for phi 1) and MSCOCO (mAP up to 0.483 for phi 4).

Maintenance & Community

The project is based on contributions from fizyr/keras-retinanet and qubvel/efficientnet. No specific community channels or active maintenance signals are mentioned.

Licensing & Compatibility

Released under the Apache License 2.0. Users should also consider the licenses of its dependencies. The Apache 2.0 license is generally permissive for commercial use.

Limitations & Caveats

The README indicates that the anchor-free version has slightly lower accuracy. Some features like quadrangle detection are detailed in separate READMEs, suggesting potential fragmentation of documentation.