We are thrilled to announce the winners of our 2021 YOLOv5 Export Competition!
With the goal of helping everyone easily train and deploy the best Vision AI models, we've organized our first Ultralytics YOLOv5 Export Competition. We value being in touch with members of our open-source community and are always impressed by numerous applications users create.
The competition ran from May 17th, 2021 to September 31st, 2021 24:00 UTC. After this date, the competition was closed and further submissions were not eligible for prize money.
The evaluation took place from September 1, 2021, to September 31, 2021. Our team went thoroughly into each submission.
The best submission ih the categories has claimed the full prize funds of $2000.00 (2000.00 USD) from Ultralytics for that category.
With the help of our amazing community, we previously created 5 categories that represent the most popular real-world deployment scenarios for YOLOv5 models, including Jetson Nano, Raspberry Pi, Google Edge TPU, Desktop CPU and Android Edge devices.
Our participants have created a public Github repository for their submission, assigned their work an open source license and posted their submission directly to one of the 5 official EXPORT Competition submissions threads to allow the community to vote. Note these threads were only for official submissions. General questions or comments have been asked directly in this thread, or in a new discussion. Links to submissions:
After much consideration, we decided on the winners for each of the five categories, which represent the most popular real-world deployment scenarios for YOLOv5 models. All participants were contacted personally and the prizes were disbursed to our winners thereafter. Today, we are happy to finally share the best solutions with you!
No winner *
No winner *
*The submissions in this category did not match the minimum set of requirements in each of the evaluation criteria. Therefore, no winner was selected for the category this time, however, there will be more chances for participants to compete again in the future.
Congratulations to the winners! Be sure to check out their repositories.
"The YOLOv5 library is great - it's updated almost daily, the models work well and the user experience is continually improving. A lot of my research involves deploying ML on embedded devices, and I'd worked with the EdgeTPU previously so this seemed like a fun challenge."
We also want to send a big shout out to everyone who participated in our Export Competition! We are fortunate to have numerous valuable members of our open-source community. It's the contributions from all of you that make our community great.
Stay amazing and keep creating! 🚀
The submissions of the Export competition were judged on the premise of several criteria: simplicity and reproducibility of their export methods, the quality of their documentation, the quality of the export and the speed and accuracy of their exported models. These submissions were then scored both by the team here at Ultralytics as well as community feedback.
The simplest export will have the least number of steps, require the least numbers of arguments/parameters, use the least number of imported packages, and be executable with the smallest amount of code.
Submissions should be well documented using either a markdown tutorial (i.e. https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data) or docs (i.e. https://docs.ultralytics.com). Each step should be explained, including setup/requirements, any settings/arguments, export steps, and deployed environment setup, if applicable.
Every aspect of export and deployment, starting from an official yolov5s.pt model, should be included. For environments that require special requirements, like Jetson Nano, all packages and/or Docker images must be supplied and documented. For Android deployments, an Android reference app should also be included. A submission must include 100% of what is required to fully export and use a YOLOv5 model.
Deployed models should return near-identical inference results to official YOLOv5 PyTorch models (i.e. inference with python detect.py --weights yolov5s.pt). Accuracy of deployed solutions will be analyzed on a hold-out test set of Ultralytics images which are unavailable to the public. Speed is very important also, with the fastest deployment solutions heavily favored. For Android, exports to GPU, NNAPI and Hexagon delegates will receive highest scoring here.
Begin your journey with the future of machine learning