The accuracy and speed are perfectly balanced, and the latest image segmentation SOTA model is released!!!

Support film and television portrait matting 、 Medical imaging analysis 、 What is the core technology behind trillion markets such as autopilot perception ？ Then we have to talk about the most important image segmentation technology . Compared with target detection 、 Image classification and other technologies , Image segmentation needs to classify each pixel , It is irreplaceable in the fine image recognition task , It is also the key to the key core competitiveness of Intelligent Vision Algorithm Engineers ！

chart 1 Image segmentation application

Because of this ,DeepLabv3、OCRNet、BiseNetv2、Fast-SCNN And other excellent algorithms emerge in endlessly , However, in the actual process of industrial implementation, it is often necessary to comprehensively consider the hardware performance 、 Accuracy and other factors , The demand for algorithms is also harsh . Often, the algorithms in the industry guarantee high recognition accuracy , It will sacrifice the running speed of the algorithm ; Instead, pursue speed , It will bring a great loss of accuracy .

chart 2 Schematic diagram of speed and accuracy balance of each algorithm

How can we balance speed and accuracy at the same time , In the current cloud 、 edge 、 Under the general industrial trend of end-to-end multi scenario collaboration, high standards meet industrial needs , It is the direction that researchers of all sessions devote themselves to .

PP-LiteSeg It is such a system that takes into account both accuracy and speed SOTA（ The best in the industry ） Semantic segmentation model . It's based on Cityscapes Data sets , stay 1080ti The upper precision is mIoU 72.0 when , The speed is as high as 273.6 FPS , （mIoU 77.5 when ,FPS by 102.6）, Beyond the existing CVPR SOTA Model STDC, The accuracy and speed are truly realized SOTA equilibrium .

chart 3 PP-LiteSeg precision / Speed description

a verbal statement without any proof , You are welcome to try it directly ！ （ Remember Star Collect and follow up the latest status ）

Portal ：

https://github.com/PaddlePaddle/PaddleSeg

What is more surprising is ,PP-LiteSeg Not only in the open source dataset, the evaluation effect is excellent , The industrial data set also shows amazing strength ！ For example, in quality inspection 、 Remote sensing scene ,PP-LiteSeg Precision and high precision 、 Big size OCRNet flat , But the speed is fast 7 times ！！！

chart 4 PP-LiteSeg and OCRNet Comparison of identification in an industrial quality inspection data set

chart 4 PP-LiteSeg and OCRNet stay deepglobe Comparison of data set identification

that PP-LiteSeg Why can it have such an excellent effect ？

PP-LiteSeg Three innovative modules are proposed ： Flexible decoding module （FLD）、 Attention fusion module （UAFM）、 Simple pyramid pool module （SPPM）.FLD Flexibly adjust the number of channels in the decoding module , Balance the calculation of encoding module and decoding module , Make the whole model more efficient ;UAFM Module effectively enhances feature representation , The accuracy of the model is better improved ;SPPM The module reduces the number of channels of the intermediate feature graph 、 Removed jump connection , Further improve the performance of the model .

chart 5 PP-LiteSeg Model structure and optimization points

It is based on the design and improvement of these modules , Final PP-LiteSeg Beyond other methods , stay 1080ti The upper precision is mIoU 72.0 when , The speed is as high as 273.6 FPS , （mIoU 77.5 when ,FPS by 102.6）, The accuracy and speed are realized SOTA Balance . More about PP-LiteSeg The content of , Please refer to ：

https://github.com/PaddlePaddle/PaddleSeg/tree/release/2.5/configs/pp_liteseg

In order to let developers have a deeper understanding PP-LiteSeg This SOTA Model , Solve the difficulties of landing application , Master the core competence of industrial practice , The flying oar team carefully prepared the excellent live broadcast class ！

Scan the code to sign up for the live class

Enter the technical exchange group

4 month 26 Japan 20:30, Baidu senior engineer will provide us with The of precision and speed balance are introduced in detail PP-LiteSeg, Disassemble its principle and use mode , There is also the actual combat of defect segmentation of automobile metal gasket , Plus live interactive Q & A , What are you waiting for ！ Hurry up and scan the code. Get in the car ！

【 References 】

chart 1

1. Auxiliary driving picture source Baidu map APP AR Navigation screenshot

2.3D The split dataset comes from MRISpineSeg spine dataset

3. Portrait cutout originates from Baidu PaddlePaddle's internal staff.

4. The remote sensing image is derived from the star map of China Science and technology GEOVIS iBrain Aerospace big data intelligent interpretation products

chart 4： Sample quality inspection data provided by partners

chart 5： From the deepglobe Data sets

END