At Edge Impulse, we’re always pushing the boundaries of what’s possible with edge AI and the last year has been no exception.
Our Applied Research team has been hard at work and, as a result, we’re excited to announce the initial release of YOLO-Pro, a new family of object detection architectures purpose-built for edge devices, accessible directly in Edge Impulse Studio. Whether you're working with high-end microcontrollers, GPUs, or accelerators, YOLO-Pro is designed to deliver top-tier performance where it matters most, on the edge.
Why did we develop YOLO-Pro?
One of the most recognized names in computer vision is YOLO — “You Only Look Once” — a family of object detection models known for their speed and accuracy. YOLO's ability to detect multiple objects in a single pass revolutionized how AI systems process visual data and made real-time detection possible. Over the years, it’s evolved through multiple iterations — referenced as YOLOv5, YOLOv8, and beyond.
These traditional YOLO models are powerful, but they were designed with cloud environments and academic datasets like COCO in mind, with restrictive licensing for commercial use. YOLO-Pro flips that paradigm. These models are engineered by Edge Impulse from the ground up to excel in real-word edge deployments, with future optimizations specifically for industrial applications currently in development. That means better performance, lower latency, and smarter resource usage for your embedded object detection projects, all available as a part of Edge Impulse without needing additional licensing.
Built for developers, by developers
This is the initial release of YOLO-Pro. As we refine the models, perform more pre-training, incorporate other improvements, and release new versions, we’re inviting our community to explore, experiment, and share feedback.
This early access allows you to:
- Test general purpose models on your edge hardware
- Influence the direction of future industrial variants
- Help shape benchmarks that reflect real-world use cases
Under the hood
The architectures and training scripts for YOLO-Pro have been written from scratch, incorporating the best innovations from classical YOLO while optimizing for the edge. The initial release contains configurable model sizes, two architecture types, built-in augmentation functionality, and the option to start model training from pre-trained weights for broad device compatibility and varied application needs.
You can find more information in our YOLO-Pro documentation.
Model sizes (parameter count)
YOLO-Pro is currently available in the following sizes:
pico(682 k)nano(2.4 M)small(6.9 M)medium(16.6 M)large(30 M)xlarge(35 M)
Architecture types
There are two options for the architecture type: Attention with SiLU and No attention with ReLU.
The attention with SiLU option is the more traditional of the two, and is similar to many modern YOLO architectures. The final block of the backbone uses partial self-attention, and all convolutional blocks use SiLU as the activation function.
The no attention with ReLU option is a custom variant developed by Edge Impulse. It has two key differences when compared to the attention with SiLU option. First, the backbone does not use partial self-attention. This choice was made because not all edge hardware is able to efficiently run the operations required in the attention layer. Secondly, all convolutional blocks use ReLU as the activation function since some edge hardware will run ReLU notably faster.
Augmentation
The data used for training the YOLO-Pro models can be augmented using both spatial transformations and color space transformations. The transformations applied are from the KerasCV preprocessing layers.
Spatial transformations change the locations of pixels but not their colors, whereas color transformations change pixel colors, but not their location. These can be used to tune augmentation based on the application. For example, a fixed camera use case may benefit from lower spatial transformations and color sensitive use cases may want no color space transformations.
Getting started with YOLO-Pro
To use YOLO-Pro in your own Edge Impulse project:
- Add an object detection learning block to your impulse.
- Select YOLO-Pro as the neural network architecture for the learning block.
- Choose your desired model size and architecture options; optionally use the
Auto configurebutton to select the model size based on your target device. - Configure your training settings.
- Train your model!
Need help with your training configuration? The EON Tuner is your best friend. It’ll guide you to the optimal choices based on your dataset and hardware.
What’s next?
As we move toward future releases, we’ll be rolling out:
- Benchmarks tailored to industrial scenarios
- Optimized variants for specific use cases
- Expanded documentation and tutorials
We’re building YOLO-Pro to be the go-to solution for edge-based object detection. Your feedback during this initial phase is invaluable. Let us know what works, what doesn’t, and what you’d love to see next.
Join us in the Edge Impulse forum, our Discord server, or tag us on social media — we’d love to see what you’re building.