DAMO-YOLO: an efficient target detection framework that takes into account both speed and accuracy

1. Introduction to target detection

The definition of target detection is to locate objects of interest in the image/space location and size.

# Generally, input an image, video or point cloud, and output the object category and detection frame coordinates. The picture on the lower left is an example of object detection on an image. There are many application scenarios for target detection, such as vehicle and pedestrian detection in autonomous driving scenarios, and berthing detection in dock management. Both of these are direct applications to object detection. Target detection is also a basic task for many CV applications, such as intrusion detection and face recognition used in factories. These require pedestrian detection and face detection as the basis to complete the detection task. It can be seen that target detection has many important applications in daily life, and its position in CV implementation is also very important, so this is a field with fierce competition.

There are currently many target detection frameworks with their own characteristics. Based on our accumulated experience in actual use, we found that the current detection framework still has the following pain points in practical application:

① Insufficient change in model scale Flexible and difficult to adapt to different computing power scenarios. For example, the detection framework of the YOLO series generally only provides the calculation amount of 3-5 models, ranging from a dozen to more than a hundred Flops, making it difficult to cover different computing power scenarios.

② The multi-scale detection capability is weak, especially the small object detection performance is poor, which makes the model application scenarios very limited. For example, in drone detection scenarios, their results are often not ideal.

③ The speed/accuracy curve is not ideal enough, and speed and accuracy are difficult to be compatible at the same time.

In response to the above situation, we designed and open sourced DAMO-YOLO. DAMO-YOLO mainly focuses on industrial implementation. Compared with other target detection frameworks, it has three obvious technical advantages:

① It integrates self-developed NAS technology and can customize models at low cost, allowing users to fully utilize the chip computing power. .

② Combining Efficient RepGFPN and HeavyNeck model design paradigms can greatly improve the multi-scale detection capabilities of the model and expand the scope of model application.

#③ Proposed a full-scale universal distillation technology that can painlessly improve the accuracy of small, medium and large models.

Below we will further analyze DAMO-YOLO from the value of 3 technical advantages.

2. DAMO-YOLO technical value

DAMO-YOLO realizes low-cost model customization based on its own The developed MAE-NAS algorithm. Models can be customized at low cost based on latency or FLOPS budget. It does not require model training or the participation of real data to provide model evaluation scores, and the model search cost is low. Targeting FLOPS can make full use of chip computing power. Searching with delay as the budget is very suitable for various scenarios that have strict requirements on delay. We also provide database construction solutions that support different hardware delay scenarios, making it easier for everyone to search using delay as a target.

The following figure shows how to use time delay for model search. First, sample the target chip or target device to obtain the delays of all possible operators, and then predict the delay of the model based on the delay data. If the predicted model magnitude meets the preset target, the model will enter subsequent model updates and score calculations. Finally, after iterative updating, the optimal model that meets the delay constraints is obtained.

Next, we will introduce how to enhance the multi-scale detection capability of the model. DAMO-YOLO combines the proposed Efficient RepGFPN and the innovative HeavyNeck, which significantly improves multi-scale detection capabilities. Efficient RepGFPN can efficiently complete multi-scale feature fusion. The HeavyNeck paradigm refers to allocating a large number of FLOPS of the model to the feature fusion layer. Such as model FLOPS ratio table. Taking DAMO-YOLO-S as an example, the calculation amount of neck accounts for nearly half of the entire model, which is significantly different from other models that mainly place the calculation amount on the backbone.

Finally introduce the distillation model. Distillation refers to transferring the knowledge of a large model to a small model, improving the performance of the small model without incurring the burden of reasoning. Model distillation is a powerful tool to improve the efficiency of detection models, but exploration in academia and industry is mostly limited to large models, and there is a lack of distillation solutions for small models. DAMO-YOLO provides a set of distillations that are common to all-scale models. This solution can not only achieve significant improvements in full-scale models, but also has high robustness. It also uses dynamic weights without needing to adjust parameters, and distillation can be completed with one-click scripts. In addition, this scheme is also robust to heterogeneous distillation, which is of great significance for the low-cost custom model mentioned above. In the NAS model, the structural similarity between the small model and the large model obtained by search is not guaranteed. If there is a heterogeneously robust distillation, the advantages of NAS and distillation can be fully exploited. The figure below shows our performance on distillation. It can be seen that no matter on the T model, S model or M model, there is a stable improvement after distillation.

##3. DAMO-YOLO application value

Based on the above technical value, how much application value can be converted? The following will introduce the comparison between DAMO-YOLO and other current SOTA detection frameworks.

DAMO-YOLO Compared with the current SOTA, the model speed is 20%-40% faster at the same accuracy, the calculation amount is reduced by 15%-50%, and the parameters are reduced by 6 %-50%, with obvious increase points in all scales and wide application range. In addition, there are obvious improvements on both small and large objects.

From the comparison of the above data, we can see that DAMO-YOLO is fast, has low Flops, and has a wide range of applications; it can also customize the model for computing power to improve the chip usage efficiency.

Relevant models have been launched on ModelScope. Inference and training can be performed through the configuration of three to five lines of code. You can experience the use. If you have any questions or problems during use, Comments are welcome in the comment area.

Next, focus on the 3 technical advantages of DAMO-YOLO and introduce the principles behind it. , to help everyone better understand and use DAMO-YOLO.

4. Introduction to the principle of DAMO-YOLO

First introduce the key technology of low-cost model customization capability MAE-NAS. Its basic idea is to regard a deep network as an information system with a continuous state space and find the entropy that can maximize the information system.

The network modeling idea is as follows: abstract the topological structure of the network F into a graph G=(V,E), where the vertex V represents the feature and the edge E represents various operators. On this basis, h(v) and h(e) can be used to represent the values ​​in vertices and edges respectively, and such a set S can be generated, which defines the continuous state space of the network, and the entropy of the set S can represent The total amount of information in the network or information system F. The information amount of the vertices measures the expressive ability of the network, and the information amount of the edges is also the entropy of the edges, which measures the complexity of the network. For the DAMO-YOLO object detection task, our main concern is to maximize the expressive ability of the network. In practical applications, only the entropy of network features is concerned. According to the Gaussian distribution differential entropy and the Gaussian entropy upper bound theorem, we use the variance of the feature map to approximate the upper bound of the network feature entropy.

In actual operation, we first initialize the weights of the network backbone with a standard Gaussian distribution, and use a standard Gaussian noise image as input. After the Gaussian noise is fed into the network for forward pass, several features can be obtained. Then the single-scale entropy, or variance, of each scale feature is calculated, and then multi-scale entropy is obtained by weighting. In the weighting process, a priori coefficients are used to balance the expressive capabilities of features at different scales. This parameter is generally set to [0,0,1,1,6]. The reason why this is set is as follows: Because in the detection model, the general features are divided into five stages, that is, five different resolutions, from 1/2 to 1/32. In order to maintain efficient feature utilization, we only utilize the last three stages. So in fact, the first two stages do not participate in the prediction of the model, so they are 0 and 0. For the other three, we have conducted extensive experiments and found that 1, 1, and 6 are a better model ratio.

Based on the above core principles, we can use the multi-scale entropy of the network as a performance proxy, Using the purification algorithm as the basic framework to search the network structure, this constitutes a complete MAE-NAS. NAS has many advantages. First of all, it supports multiple inference budget restrictions, and can use FLOPS, parameter amount, latency and network layer number to conduct a model search. Secondly, it also supports a very large number of variations in fine-grained network structures. Because evolutionary algorithms are used here to perform network searches, the more variants of network structures are supported, the higher the degree of customization and flexibility during search. In addition, in order to facilitate users to customize the search process, we provide official tutorials. Finally, and most importantly, MAE-NAS is zero-short, that is, its search does not require any actual data participation and does not require any actual model training. It searches for tens of minutes on the CPU and can produce an optimal network result under the current constraints.

In DAMO-YOLO, we use MAE-NAS to search the backbone network of T/S/M model with different delays as search targets; The backbone network infrastructure is packaged, small models use ResStyle, and large models use CSPStyle.

As can be seen from the table below, CSP-Darknet is a manually designed network using the CSP structure. It has also achieved some results in YOLO v 5 /V6 Wide range of applications. We used MAE-NAS to generate a basic structure, and after packaging it with CSP, we found that the model was significantly improved in speed and accuracy. In addition, you can see the MAE-ResNet form on small models, which will have higher accuracy. There is a clear advantage in using the CPS structure on large models, which can reach 48.7.

How to use MAE-NAS to search backbone? Here we introduce our TinyNAS toolbox, which is already online in ModelScope. You can easily get the desired model through visual configuration on the web page. At the same time, MAE-NAS has also been open sourced on github. Interested students can search for the desired model with greater freedom based on the open source code.

Next, we will introduce how DAMO-YOLO improves multi-scale detection capabilities, which relies on the fusion of different scale features of the network. In previous detection networks, the depth of features at different scales varies greatly. For example, large-resolution features are used to detect small objects, but their feature depth is shallow, which will affect the small object detection performance.

A work we proposed at ICLR2022 - GFPN, processes high-level semantic information and low-level spatial information at the same time with the same priority, and is very friendly to the fusion and complementation of multi-scale features. In the design of GFPN, we first introduced a skip layer in order to enable GFPN to be designed deeper. We use a log2n-link to reuse features and reduce redundancy.

Queen fusion is to increase the interactive fusion of features of different scales and features of different depths. In addition to receiving different scale features diagonally above and below it, each node in Queen fusion also receives different scale features at the same feature depth, which greatly increases the amount of information during feature fusion and promotes multi-scale information at the same depth. fusion on.

Although GFPN’s feature reuse and unique connection design have brought improvements in model accuracy. Since our skip layer and our Queen fusion bring fusion operations on multi-scale feature nodes, as well as upsampling and downsampling operations, they greatly increase the time-consuming inference and make it difficult to meet the implementation requirements of the industry. So in fact, GFPN is a FLOPS efficient, but delay inefficient structure. In view of some defects of GFPN, we analyzed and attributed the reasons as follows:

① First of all, features of different scales actually share the number of channels, and there are many Features are redundant and network configuration is not flexible enough.

② Second, there are upsampling and downsampling connections in the Queen feature, and the time consumption of the upsampling and downsampling operators is significantly improved.

③ Third, when nodes are stacked, serial connections at the same feature depth reduce the parallel efficiency of the GPU, and each stack brings The growth of serial paths has been significant.

#To address these problems, we have made corresponding optimizations and proposed Efficient RepGFPN.

##In optimization, it is mainly divided into two categories. One is the optimization of topology structure. , the other category is the optimization of fusion methods.

In terms of topology optimization, Efficient RepGFPN uses different channel numbers for different scale features, so that it can flexibly control high-level features and low-level features under the constraints of lightweight calculations. expressive ability. In the case of FLOPS and delay approximation, flexible configuration can achieve the best accuracy and speed efficiency. In addition, we also conducted an efficiency analysis on a connection in queen fusion and found that the upsampling operator has a huge burden, but the accuracy improvement is small, which is far lower than the benefit of the downsampling operator. So we removed the upsampling connection in the queen fusion. As can be seen in the table, the ticks diagonally downward are actually upsampling, and the ticks diagonally upward are downsampling. You can compare it with the picture on the left to see that small resolutions gradually become larger resolutions downwards, and the connections to the lower right represent The purpose is to upsample small-resolution features, connect them to large-resolution features, and fuse them into large-resolution features. The final conclusion is that the downsampling operator has higher returns, while the upsampling operator has very low returns, so we removed the upsampling connection in the Queen feature to improve the efficiency of the entire GFPN.

In terms of integration methods, we have also made some optimizations. First, fix the number of fusion nodes so that only two fusions are performed in each model instead of continuously stacking fusions to create a deeper GFPN as before. This avoids the parallel efficiency caused by the continuous growth of serial links. reduce. In addition, we specially designed a fusion block for feature fusion. In fusion block, we introduce technologies such as heavy parameterization mechanism and multi-layer aggregation connection to further improve the fusion effect.

In addition to Neck, the detection head Head is also an important part of the detection model. It takes the features output by Neck as input and is responsible for outputting regression and classification results. We designed experiments to verify the trade off between Efficient RepGFPN and Head, and found that when the model latency is strictly controlled, the deeper the Efficient RepGFPN, the better. Therefore, in the network design, the calculation amount is mainly allocated to Efficient RepGFPN, while only one layer of linear projection is reserved in the Head part for classification and regression tasks. We call the Head that has only one layer of classification and one layer of regression non-linear mapping layer ZeroHead. A design pattern that allocates this computational load mainly to Neck is called the HeavyNeck paradigm.

The final model structure of DAMO-YOLO is shown in the figure below.

The above are some thoughts on in model design. Finally, let’s introduce the distillation scheme.

Take the output features of Efficient RepGFPN from DAMO-YOLO for distillation. The student feature will first pass through the alignmodule to align its channel number to the teacher. In order to remove the bias of the model itself, the features of the student and teacher will be normalized by unbiased BN, and then the distillation loss calculation will be performed. During distillation, we observed that excessive loss will hinder the convergence of the student's own classification branch. So we chose to use a dynamic weight that decays with training. From the experimental results, the dynamic uniform distillation weight is robust to T/S/M models.

The distillation chain of DAMO-YOLO is, L distillation M, M distillation S. It is worth mentioning that when M distills S, M uses CSP packaging, while S uses Res packaging. Structurally speaking, M and S are isomers. However, when using the DAMO-YOLO distillation scheme, M distills S, there can also be an improvement of 1.2 points after distillation, indicating that our distillation scheme is also robust to isomerism. So in summary, DAMO-YOLO’s distillation scheme has free parameters, supports a full range of models, and is heterogeneous and robust.

##Finally, let’s summarize DAMO-YOLO . DAMO-YOLO combines MAE-NAS technology to enable low-cost model customization and fully utilizes chip computing power. Combined with Efficient RepGFPN and HeavyNeck paradigms, it improves multi-scale detection capabilities and has a wide range of model applications. With the full-scale distillation scheme, it can Further improve model efficiency.

The DAMO-YOLO model has been launched on ModelScope and is open source on github. Everyone is welcome to try it.

5. DAMO-YOLO Development Plan

DAMO-YOLO has just been released, and there are still many things that need to be improved and improved Optimization place. We plan to improve the deployment tools and support ModelScope in the short term. In addition, more application examples will be provided based on the competition champion solutions within the group, such as UAV small target detection and rotating target detection. It is also planned to launch more example models, including the Nano model for the device and the Large model for the cloud. Finally, I hope everyone will pay attention and provide positive feedback.

The above is the detailed content of DAMO-YOLO: an efficient target detection framework that takes into account both speed and accuracy. For more information, please follow other related articles on the PHP Chinese website!