YOLO-Pro Does More With Less

Do you know YOLO? If you are doing any work with object detection models, you should. It is one of the most prominent real-time object detection algorithms available in the world of computer vision today. It is known for its state-of-the-art performance and utility in a wide range of applications. But while YOLO is capable of processing images in a single forward pass of a neural network, it is still not efficient enough for use with many of the tiny hardware platforms commonly used in edge AI.

Or at least, it wasn’t efficient enough in the past. But now the team over at Edge Impulse has shrunk YOLO models down in a way that maintains high levels of accuracy, while allowing them to run even on low-power computing platforms. Called YOLO-Pro, these custom models range in size from 682K to 35M parameters, and they have some other optimizations that allow for ideal operation on everything from microcontrollers to Particle Tachyons and NVIDIA Jetsons.

Edge Impulse’s Applied Research team designed YOLO-Pro from the ground up to meet the unique demands of edge computing, where power, memory, and latency constraints make traditional cloud-trained AI models impractical. While the original YOLO family of models forever changed object detection with its speed and ability to identify multiple objects in a single pass, they were developed with large-scale computing environments and academic datasets in mind. YOLO-Pro turns that model on its head by being purpose-built for deployment on embedded devices.

Another one of YOLO-Pro’s biggest advantages is its accessibility. Traditional YOLO models often come with restrictive licenses that can complicate commercial use. Edge Impulse’s YOLO-Pro models, on the other hand, which are fully integrated into Edge Impulse Studio, allow developers to train, deploy, and test models directly without any additional licensing barriers.

Developers have flexibility in how they tailor YOLO-Pro to their needs. The models are available in six sizes — from “pico” (682K parameters) to “xlarge” (35M parameters) — and offer two architecture options: "Attention with SiLU" for more traditional YOLO-like performance, and "No attention with ReLU", an Edge Impulse-developed variant optimized for devices that struggle with attention layers or that run ReLU operations faster.

Training is further enhanced with robust data augmentation capabilities, including both spatial and color transformations, allowing users to fine-tune their datasets for specific use cases, like fixed cameras or color-sensitive applications.

Getting started involves adding an object detection learning block in Edge Impulse Studio and selecting YOLO-Pro as the architecture. The built-in EON Tuner then assists in automatically selecting the best model configuration for the target hardware.

This is just the first release of YOLO-Pro, and Edge Impulse plans to expand it with industrial benchmarks, specialized model variants, and richer documentation in future updates. Be sure to check out the documentation if you want to integrate these new models into your projects.