EfficientPose provides an end-to-end 6D multi-object pose estimation approach, building upon EfficientDet and RetinaNet. It is designed for researchers and developers working on computer vision tasks requiring accurate and efficient object pose estimation, particularly in industrial or robotics applications.

How It Works

This project implements a novel approach to 6D pose estimation by leveraging the efficient architecture of EfficientDet. It integrates feature extraction, bounding box prediction, and keypoint estimation into a single, scalable network. The architecture is designed to be computationally efficient without sacrificing accuracy, making it suitable for real-time applications.

Quick Start & Requirements

• Installation: Clone the repository, create a conda environment (conda create -n EfficientPose python==3.7), activate it (conda activate EfficientPose), install TensorFlow 1.15.0 (conda install tensorflow-gpu==1.15.0), and then install dependencies (pip install -r requirements.txt). Compile Cython modules with python setup.py build_ext --inplace.
• Prerequisites: Python 3.7, TensorFlow 1.15.0 (GPU recommended), Cython.
• Datasets & Weights: Pretrained weights and Linemod/Occlusion datasets are available for download.
• Links: Official Implementation

Highlighted Details

• End-to-end 6D multi-object pose estimation.
• Based on EfficientDet and RetinaNet architectures.
• Supports Linemod and Occlusion datasets.
• Includes scripts for training, evaluation, inference (directory and webcam), benchmarking, and debugging.

Maintenance & Community

The project is authored by Yannick Bukschat and Marcus Vetter. Further community engagement details are not explicitly provided in the README.

Licensing & Compatibility

• License: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).
• Compatibility: Freely available for non-commercial use. Commercial use requires contacting the authors.

Limitations & Caveats

The README notes that the Linemod and Occlusion datasets are small and may not yield reasonable real-world pose estimation performance. Users need to replace default camera parameters with their specific webcam parameters for live inference.