Unleashing the Power of 9 with YOLOv9

Official GitHub Repo: WongKinYiu/yolov9: Implementation of paper - YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information (github.com)

?? Introduction ??

Microsoft and Apple might've missed the Digit 9 while releasing the products, but team of Chien-Yao Wang, I-Hau Yeh, and Hong-Yuan Mark Liao made sure that the world comes in face with YOLOv9 models. This latest addition to the You Only Live Once (or YOLO) family brings forth novel concepts, including the introduction of Generalized ELAN (GELAN) and Integrated PGI, propelling computer vision into uncharted territory. But fear not, fellow enthusiasts, for we're here to unravel the mysteries of YOLOv9, one exciting step at a time! Yep, some terms might be a little heavy, but let's explore the YOLOv9 model and its history slowly, while taking one step at a time.

?? Background ??

Initially developed by Joseph Redmon et. al. in 2015 [Read Original Paper Here], YOLO models took the world by storm. The usage of a single neural network for both classification of image and objects, generating class probabilities and also the bounding box locations were some new approaches which overtook previous models like DPM and R-CNN (At that time ofc). The later additions such as YOLOv2 (2016) and YOLOv3 (2018) improved on the foundation laid by the original model, with increased accuracy and predictions. Alongside this, the addition of new layers such as normalization and feature pyramid networks promoted the growth and complexity of the base model.

The further refinements over v2 and v3, such as introduction of new terms like CSPDarknet and PANet, configuration of a single model into 3 parts - head, tail and backbone [Read Yolov3 Here], enhanced the computational efficiency and power of the model, leading to v4 of the model. YOLOv4 (2020) was succeeded by YOLOv5 (2020) with quicker training and inference time. At this point, YOLO makers also decided to make different variants for facilitating different tasks on various devices - nano, mini, small, medium, large (feels like I might miss something)(Maybe some mini/hybrid models are left).

Decoupling a YOLOv5 model and reorganizing it into a new form with newer tech lead to a YOLOv6 (2022) model, with increased performance. The size of YOLO models at this point exceeded the expectations of makers and required a little rework. Introduction of E-ELAN into the model was the solution of this - building a YOLOv7 (2022) model.

So now, we have 7 models. The final model before the 9th variant - YOLOv8 (2023), comes with some more refined features. The v8 model features an anchor free model. This made it interesting for researchers to use a new refined YOLO model.

Now the then history is complete, let's move towards the new kid in the town - YOLOv9

?? Understanding YOLOv9 ??

YOLOv9 is the epitome of progress, innovation, and real-time object detection. With the introduction of cutting-edge features like Programmable Gradient Information (PGI) and Generalized Efficient Layer Aggregation Network (GELAN), YOLOv9 propels the realm of computer vision into uncharted territory. These innovations pave the way for unparalleled object detection capabilities, setting a new benchmark for excellence in the field.

>> (Feb 2024) As of writing this, YOLOv9 only supports object-detection. Segmentation, classification, and other task types are not supported at this time. [Read More Here]

? What is Programmable Gradient Information (PGI)? ?

PGI is not just another buzzword; it's a game-changer! The main objective behind introducing PGI is to retain as much information as possible, which comes in handy when calculating the objective function. In simple terms, it refers to capturing gradient information. As the model moves forward in the network, previous information which can be crucial may be lost if it occurs in early layers. As a result, PGI will help in securing this gradient information, allowing for better results. This indirectly helps in tackling information bottleneck occurring in some previous models.

??Understood PGI, but what is this new GELAN???

Generalized Efficient Layer Aggregation Network (GELAN) may sound like a mouthful, but its impact speaks volumes. GELAN [Pretty heavy Term] is basically a lightweight neural network architecture. This architecture allows even the simplest of convolutional nets to maximize the parameter efficiency. Also, it provides a very stable performance across various computational blocks of different depth and widths. The main focus of GELAN's is to improve the object detection of any and every model.

?? Meet the Family: YOLOv9 Variants ??

YOLOv9 doesn't believe in a one-size-fits-all approach. With four distinct variants - v9-S, v9-M, v9-C, and v9-E - this powerhouse of a model offers versatility and adaptability like never before. Whether you're working with different devices, configurations, or use cases, there's a YOLOv9 variant tailored just for you!

The details and weights of S-M models are not available at the time of writing. Make sure to read the official research paper of YOLOv9 (Learn More About YOLOv9). Let's wait till we have more information about the models before concluding anything.

?? Features to Anticipate ??

• Enhanced Accuracy: The above 2 architectures, combined with better training strategies and added support for TPUs (Yes, Colab users are going to be very happy)
• Adaptability and Versatility: The newer model now supports even more categories and can be used for multiple use cases, backed by easy integration.
• Seamless Integration: Allows the YOLOv9 to be integrated and combined with other pre-trained/self-developed models.
• Better performance than current SOTA models: In contrast to YOLOv7 AF, YOLOv9-C achieves a reduction of 42% in parameters and 21% in calculations, while still preserving a consistent 53% accuracy in Average Precision (AP). When compared to YOLOv8-X, YOLOv9-X boasts a reduction of 15% in parameters and 25% in calculations, accompanied by a notable enhancement of 1.7% in AP. (Can be found here)

All the performance metrics are based by training them on MS COCO Dataset, where all the models were compared on same metrics and parameters.

??So, how do we use the latest v9 for our tasks? (Conclusion) ??

The implementation part of this would v9 model will be shared in the next article, after everything is clear from my side and understandable. But right now, I think people might've understand about YOLOv9 model and its impact, new features and what not. So, that's it for this article. Stay tuned for the next one. Thanks for reading!!

#YOLOv9 #ComputerVision #ObjectDetection #ArtificialIntelligence #MachineLearning #Innovation #Technology #DeepLearning #NeuralNetworks #AI #YOLO #CV