YOLOX: Powerful Object Detection Meets a Maintenance Challenge#
YOLOX (You Only Look Once - eXceeding), introduced by Megvii Technology in 2021, quickly established itself as a high-performance model in the popular YOLO family of real-time object detectors. Its key innovations, including an anchor-free design and a decoupled head architecture, offered state-of-the-art performance (mAP) across various model sizes (from YOLOX-Nano to YOLOX-X) while maintaining competitive inference speeds (FPS).
However, despite its strengths and permissive Apache 2.0 license, the original YOLOX repository hasn't seen significant updates since 2022. As the computer vision community noted (including discussions on platforms like Reddit), this led to increasing friction for users: compatibility issues arose with modern Python versions, core dependencies like PyTorch became outdated, and environments like Google Colab presented challenges. Effectively using YOLOX became difficult, particularly compared to continuously updated alternatives.
There was a clear community need for a readily usable, actively maintained version that preserved the valuable Apache 2.0 license.
Pixeltable Steps In: Introducing pixeltable-yolox#
As developers of Pixeltable, a multimodal data store for AI workloads, we frequently integrate cutting-edge models. We encountered these YOLOX usability challenges firsthand and recognized the broader community need.
We are excited to introduce pixeltable-yolox, our fork of the original YOLOX library. Our primary goal is not to fundamentally change YOLOX's core models but to make the existing, powerful feature set accessible, reliable, and easy to integrate for everyone in the CV community, under the original permissive license.
What's New in pixeltable-yolox? Focused on Usability & Compatibility#
Based on community feedback and our own integration needs, we've focused on modernization and ease of use:
- Modern Python Compatibility: Easily
pip install-able and compatible with current Python versions (3.9+). - Updated Dependencies: Works seamlessly with recent versions of PyTorch and other essential libraries.
- Simplified Inference API: Introduced a new
YoloxProcessorclass, inspired by the Hugging Face Transformers API, for cleaner pre/post-processing separation. - Refactored CLI: Improved the command-line interface for training and evaluation workflows.
- Enhanced Testing: Added comprehensive tests for better code reliability.
Ongoing work includes CI testing across environments and adding type hinting for MyPy compatibility.
Getting Started with the Maintained Fork#
Using pixeltable-yolox is designed to be straightforward:
Installation:
Inference Example (Python):
For training details and advanced usage, please refer to the repository README.
Our Commitment: Stability under Apache 2.0#
Pixeltable is committed to maintaining this fork for the foreseeable future, ensuring it remains compatible with current environments and dependencies. Crucially, all improvements are provided under the original, permissive Apache 2.0 license.
Our focus is on the stability and usability of the *existing* YOLOX models and features. While we don't plan major new architectural enhancements ourselves, we warmly welcome community contributions! We provide the engineering infrastructure (CI, releases) and guidance. See our contributors' guide.
Join the Effort!#
We believe this fork addresses a genuine need for a reliable, Apache-licensed, anchor-free object detector. We invite you to:
- Try it out:
pip install pixeltable-yolox - Give Feedback: Report issues or suggest improvements on the GitHub repository.
- Contribute: Help us keep YOLOX a valuable asset for everyone.
You can follow the journey and discussion on the Reddit update thread as well.
In Memory of Dr. Jian Sun#
We want to echo the sentiment in the original YOLOX repository and acknowledge the foundational work guided by Dr. Jian Sun. His contributions were instrumental in YOLOX being released and open-sourced, and his passing is a significant loss to the computer vision field.
