YOLO is immortal! YOLOv9 is released: performance and speed SOTA~

Today’s deep learning methods focus on designing the most suitable objective function so that the model’s prediction results are closest to the actual situation. At the same time, a suitable architecture must be designed to obtain sufficient information for prediction. Existing methods ignore the fact that when the input data undergoes layer-by-layer feature extraction and spatial transformation, a large amount of information will be lost. This article will delve into important issues when transmitting data through deep networks, namely information bottlenecks and reversible functions. Based on this, the concept of programmable gradient information (PGI) is proposed to cope with the various changes required by deep networks to achieve multi-objectives. PGI can provide complete input information for the target task to calculate the objective function, thereby obtaining reliable gradient information to update network weights. In addition, a new lightweight network architecture-Generalized Efficient Layer Aggregation Network (GELAN) based on gradient path planning is designed.

The verification results show that the GELAN architecture has gained significant advantages through PGI on lightweight models. Experiments on the MS COCO data set show that GELAN combined with PGI can achieve better parameter utilization than the state-of-the-art methods based on deep convolution using only traditional convolution operators. PGI's versatility makes it suitable for a wide range of models, from lightweight to large models. With PGI, the model is fully informed, so it is possible to achieve better results using a model trained from scratch than a state-of-the-art model pre-trained on a large dataset.

Article address: https://arxiv.org/pdf/2402.13616

Code link: https://github.com/WongKinYiu/yolov9

Excellent Performance

According to the comparison results of real-time target detectors on the MS COCO dataset, the target detection methods based on GELAN and PGI are significantly ahead of the previous methods trained from scratch in terms of target detection performance. The new method outperforms RT DETR, which relies on large dataset pre-training, in terms of accuracy, and also outperforms YOLO MS based on deep convolution design in terms of parameter utilization. These results indicate that GELAN and PGI methods have potential advantages in the field of target detection and may become important technology choices in future research and applications.

Contributions of this article

1. The existing deep neural network architecture is theoretically analyzed from the perspective of reversible functions. Through this This process has successfully explained many phenomena that were difficult to explain in the past. PGI and auxiliary reversible branches were also designed based on this analysis and achieved excellent results. The PGI designed by
2. solves the problem that deep supervision can only be used for extremely deep neural network architectures, making the new lightweight architecture truly applicable to daily work.
3. The designed GELAN uses only traditional convolutions to achieve higher parameter usage than deep convolution designs based on state-of-the-art technology, while showing great advantages of being lightweight, fast and accurate.
4. Combining the proposed PGI and GELAN, the object detection performance of YOLOv9 on the MS COCO dataset greatly exceeds existing real-time object detectors in all aspects.

Method

PGI and related network architecture and methods

As shown in the figure below, (a) Path Aggregation Network (PAN), (b) Reversible Column (RevCol), (c) traditional deep supervision, and (d) Programmable Gradient Information (PGI) proposed by YOLOv9.

PGI is mainly composed of three components:

1. Main branch: the architecture used for inference;
2. Auxiliary reversible Branch: Generate reliable gradients for the main branch to transmit backward;
3. Multi-level auxiliary information: Control the main branch to learn planable multi-level semantic information.

The architecture of GELAN

As shown in the figure below, (a) CSPNet, (b) ELAN, and (c) GELAN proposed by YOLOv9. It imitates CSPNet and extends ELAN to GELAN, which can support any computing block.

Comparison of results

Comparison with existing technology

The following table lists A comparison of YOLOv9 with other real-time object detectors trained from scratch is presented. Overall, the best performing methods among existing methods are YOLO MS-S for lightweight models, YOLO MS for medium models, YOLOv7 AF for general models, and YOLOv8-X for large models. Compared with YOLO MS of lightweight and medium models, YOLOv9 has about 10% fewer parameters and 5∼15% less calculations, but still has a 0.4∼0.6% improvement in AP. Compared with YOLOv7 AF, YOLOv9-C has 42% fewer parameters and 21% fewer calculations, but achieves the same AP (53%). Compared with YOLOv8-X, YOLOv9-X has 15% fewer parameters, 25% fewer calculations, and a significant improvement in AP, which has increased by 1.7%. The above comparison results show that YOLOv9 is significantly improved over existing methods in all aspects.

Comparison with the most advanced real-time object detectors

The methods participating in the comparison all use ImageNet as pre-training weights, including RT DETR, RTMDet and PP-YOLOE, etc. YOLOv9 using the scratch training method clearly exceeds the performance of other methods.

Visualized results

Feature map (visualized results): By PlainNet , Random initial weight output of ResNet, CSPNet and GELAN at different depths. After 100 layers, ResNet starts to produce feed-forward output that is enough to confuse the target information. The GELAN proposed here can still retain quite complete information at the 150th layer, and still has sufficient discrimination ability at the 200th layer.

PAN feature maps (visualization results) of GELAN and YOLOv9 (GELAN PGI): After a round of bias warm-up. GELAN had some initial divergence, but after adding the reversible branch of PGI, it was better able to focus on the target object.

Visualized results of random initial weight output feature maps of different network architectures: (a) Input image, (b) PlainNet, (c) ResNet, (d) CSPNet and (e) proposed GELAN. As can be seen from the figure, in different architectures, the degree of information provided to calculate the loss of the objective function is different, and our architecture can retain the most complete information and provide the most reliable gradient information for calculating the objective function.

Conclusion

This paper proposes to use PGI to solve the problem of information bottlenecks and deep supervision mechanisms that are not suitable for lightweight neural networks. question. designed GELAN, an efficient and lightweight neural network. In terms of target detection, GELAN shows strong and stable performance under different computing modules and depth settings. It is indeed broadly scalable to models suitable for a variety of inference devices. In response to the above two problems, the introduction of PGI enables both lightweight models and deep models to achieve significant improvements in accuracy. YOLOv9, designed by combining PGI and GELAN, shows strong competitiveness. Its excellent design allows the deep model to reduce the number of parameters by 49% and the calculation amount by 43% compared with YOLOv8, but still achieves a 0.6% AP improvement on the MS COCO data set.

Original link: https://mp.weixin.qq.com/s/nP4JzVwn1S-MeKAzbf97uw

The above is the detailed content of YOLO is immortal! YOLOv9 is released: performance and speed SOTA~. For more information, please follow other related articles on the PHP Chinese website!