The latest installation in the YOLO series, YOLOv9, was released on February 21st, 2024. Since its inception in 2015, the YOLO (You Only Look Once) object-detection algorithm has been closely followed by tech enthusiasts, data scientists, ML engineers, and more, gaining a massive following due to its open-source nature and community contributions. With every new release, the YOLO architecture becomes easier to use and much faster, lowering the barriers to use for people around the world.
YOLO was introduced as a research paper by J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, signifying a step forward in the real-time object detection space, outperforming its predecessor – the Region-based Convolutional Neural Network (R-CNN). It is a single-pass algorithm having only one neural network to predict bounding boxes and class probabilities using a full image as input.
What is YOLOv9?
YOLOv9 is the latest version of YOLO, released in February 2024, by Chien-Yao Wang, I-Hau Yeh, and Hong-Yuan Mark Liao. It is an improved real-time object detection model that aims to surpass all convolution-based, and transformer-based methods.
YOLOv9 is released in four models, ordered by parameter count: v9-S, v9-M, v9-C, and v9-E. To improve accuracy, it introduces programmable gradient information (PGI) and the Generalized Efficient Layer Aggregation Network (GELAN). PGI prevents data loss and ensures accurate gradient updates and GELAN optimizes lightweight models with gradient path planning.
At this time, the only computer vision task supported by YOLOv9 is object detection.
YOLO Version History
Before diving into the YOLOv9 specifics, let’s briefly recap on the other YOLO versions available today.
YOLOv1
YOLOv1 architecture (displayed above) surpassed R-CNN with a mean average precision (mAP) of 63.4, and an inference speed of 45 FPS on the open-source Pascal VOC 2007 dataset. With YOLOv1, object detection is treated as a regression task to predict bounding boxes and class probabilities from a single pass of an image.
YOLOv2
Released in 2016, it could detect 9000+ object categories. YOLOv2 introduced anchor boxes – predefined bounding boxes called priors that the model uses to pin down the ideal position of an object. YOLOv2 achieved 76.8 mAP at 67 FPS on the VOC 2007 dataset.
YOLOv3
The authors released YOLOv3 in 2018 which boasted higher accuracy than previous versions, with an mAP of 28.2 at 22 milliseconds. To predict classes, the YOLOv3 model uses Darknet-53 as the backbone with logistic classifiers instead of softmax and Binary Cross-entropy (BCE) loss.
YOLOv4
2020, Alexey Bochkovskiy et al. released YOLOv4, introducing the concept of a Bag of Freebies (BoF) and a Bag of Specials (BoS). BoF is a set of data augmentation techniques that increase accuracy at no additional inference cost. (BoS significantly enhances accuracy with a slight increase in cost). The model achieved 43.5 mAP at 65 FPS on the COCO dataset.
YOLOv5
Without an official research paper, Ultralytics released YOLOv5 also in 2020. The model is easy to train since it is implemented in PyTorch. The model architecture uses a Cross-stage Partial (CSP) Connection block as the backbone for a better gradient flow to reduce computational cost. YOLOv5 uses YAML files instead of CFG files in the model configurations.
YOLOv6
YOLOv6 is another unofficial version introduced in 2022 by Meituan – a Chinese shopping platform. The company targeted the model for industrial applications with better performance than its predecessor. The changes resulted in YOLOv6n achieving an mAP of 37.5 at 1187 FPS on the COCO dataset and YOLOv6s achieving 45 mAP at 484 FPS.
YOLOv7
In July 2022, a group of researchers released the open-source model YOLOv7, the fastest and the most accurate object detector with an mAP of 56.8% at FPS ranging from 5 to 160. YOLOv7 is based on the Extended Efficient Layer Aggregation Network (E-ELAN), which improves training by letting the model learn diverse features with efficient computation.
YOLOv8
YOLOv8 has no official paper (as with YOLOv5 and v6) but boasts higher accuracy and faster speed for state-of-the-art performance. For instance, the YOLOv8m has a 50.2 mAP score at 1.83 milliseconds on the MS COCO dataset and A100 TensorRT. YOLO v8 also features a Python package and CLI-based implementation, making it easy to use and develop.
Since YOLOv9’s February 2024 release, another team of researchers has released YOLOv10 (May 2024), for real-time object detection.
Architecture YOLOv9
To address the information bottleneck (data loss in the feed-forward process), YOLOv9 creators propose a new concept, i.e. the programmable gradient information (PGI). The model generates reliable gradients via an auxiliary reversible branch. Deep features still execute the target task and the auxiliary branch avoids the semantic loss due to multi-path features.
The authors achieved the best training results by applying PGI propagation at different semantic levels. The reversible architecture of PGI is built on the auxiliary branch, so there is no additional cost. Since PGI can freely select a loss function suitable for the target task, it also overcomes the problems encountered by mask modeling.
The proposed PGI mechanism can be applied to deep neural networks of various sizes. In the paper, the authors designed a generalized ELAN (GELAN) that simultaneously takes into account the number of parameters, computational complexity, accuracy, and inference speed. The design allows users to choose appropriate computational blocks arbitrarily for different inference devices.
Using the proposed PGI and GELAN – the authors designed YOLOv9. To conduct experiments they used the MS COCO dataset, and the experimental results verified that the proposed YOLO v9 achieved the top performance in all cases.
Research Contributions
- Theoretical analysis of deep neural network architecture from the perspective of reversible function. The authors designed PGI and auxiliary reversible branches based on this analysis and achieved excellent results.
- The designed PGI solves the problem that deep supervision can only be used for extremely deep neural network architectures. Thus, it allows new lightweight architectures to be truly applied in daily life.
- The GELAN network only uses conventional convolution to achieve a higher parameter usage than the depth wise convolution design. So it shows the great advantages of being light, fast, and accurate.
- Combining the proposed PGI and GELAN, the object detection performance of the YOLOv9 on the MS COCO dataset largely surpasses the existing real-time object detectors in all aspects.
YOLOv9 License
YOLOv9 was not released with an official license. In the following days, however WongKinYiu updated the official license to GPL-3.0. YOLOv7 and YOLOv9 have been released under WongKinYiu’s repository.
Advantages of YOLOv9
YOLOv9 arises as a powerful model, offering innovative features that will play an important role in the further development of object detection, and maybe even image segmentation and classification down the road. It provides faster, clearer, and more flexible actions, and other advantages include:
- Handling the information bottleneck and adapting deep supervision to lightweight architectures of neural networks by introducing the Programmable Gradient Information (PGI).
- Creating the GELAN, a practical and effective neural network. GELAN has proven its strong and stable performance in object detection tasks at different convolution and depth settings. It could be widely accepted as a model suitable for various inference configurations.
- By combining PGI and GELAN – YOLOv9 has shown strong competitiveness. Its clever design allows the deep model to reduce the number of parameters by 49% and the number of calculations by 43% compared with YOLOv9. And it still has a 0.6% Average Precision improvement on the MS COCO dataset.
- The developed YOLOv9 model is superior to RT-DETR and YOLO-MS in terms of accuracy and efficiency. It sets new standards in lightweight model performance by applying conventional convolution for better parameter utilization.
| Model | #Param. | FLOPs | AP50:95val | APSval | APMval | APLval |
|---|---|---|---|---|---|---|
| YOLOv7 [63] | 36.9 | 104.7 | 51.2% | 31.8% | 55.5% | 65.0% |
| + AF [63] | 43.6 | 130.5 | 53.0% | 35.8% | 58.7% | 68.9% |
| + GELAN | 41.7 | 127.9 | 53.2% | 36.2% | 58.5% | 69.9% |
| + DHLC [34] | 58.1 | 192.5 | 55.0% | 38.0% | 60.6% | 70.9% |
| + PGI | 58.1 | 192.5 | 55.6% | 40.2% | 61.0% | 71.4% |
The above table demonstrates average precision (AP) of various object detection models.
YOLOv9 Applications
YOLOv9 is a flexible computer vision model that you can use in different real-world applications. Here we suggest a few popular use cases.
- Logistics and distribution: Object detection can assist in estimating product inventory levels to ensure sufficient stock levels and provide information regarding consumer behavior.
- Autonomous vehicles: Autonomous vehicles can utilize YOLOv9 object detection to help navigate self-driving cars safely through the road.
- People counting: Retailers and shopping malls can train the model to detect real-time foot traffic in their shops, detect queue length, and more.
- Sports analytics: Analysts can use the model to track player movements in a sports field to gather relevant insights regarding team performance.
YOLOv9: Main Takeaways
The YOLO models are the standard in the object detection space with their great performance and wide applicability. Here are our first conclusions about YOLOv9:
- Ease-of-use: YOLOv9 is already in GitHub, so the users can implement YOLOv9 quickly through the CLI and Python IDE.
- YOLOv9 tasks: YOLOv9 is efficient for real-time object detection with improved accuracy and speed.
- YOLOv9 improvements: YOLOv9’s main improvements include a decoupled head with anchor-free detection and mosaic data augmentation that turns off in the last ten training epochs.
In the future, we look forward to seeing if the creators will expand YOLOv9 capabilities to a wide range of other computer vision tasks as well.
Viso Suite is the end-to-end platform for computer vision. Viso Suite offers a host of pre-trained models to choose from, or the possibility to import or train your own custom AI models. To learn how you can solve your industry’s challenges with computer vision, book a demo of Viso Suite.
