EMOv2:推进5M视觉模型前沿
概括
本研究介绍了Efficient MOdel版本2 (EMOv2),这是一个新的轻量级神经网络架构. EMOv2为约500万个参数的模型实现了最新的性能,为移动应用程序推进了高效的深度学习.
科学领域:
- 计算机视觉 计算机视觉
- 机器学习 机器学习
- 深度学习 (Deep Learning) 是一种深度学习.
背景情况:
- 轻量级模型对于在资源有限的移动设备上有效部署至关重要.
- 现有的轻量级卷积神经网络 (CNN) 架构,如反向剩余块 (IRB),缺乏基于注意力的对应.
- 变压器提供了注意力机制,但对于轻量级应用程序来说,它可能是计算密集的.
研究的目的:
- 为密集的预测任务开发参数效率和轻量级模型.
- 为500万参数大小范围内的模型建立一个新的性能边界.
- 为了统一高效的IRB和变压器组件的设计原则.
主要方法:
- 通过统一高效的IRB和变压器组件,重新设计了轻量级基础设施.
- 抽象一个剩余的Meta移动区块 (MMBlock) 用于轻量化模型设计.
- 开发了一个改进的倒置剩余移动块 (i222222222222222222222222rmb) 和一个层次化的高效模型版本2 (EMOv2).
主要成果:
- EMOv2模型 (1M,2M,5M参数) 实现了72.3,75.8和79.4的Top-1精度,明显超过了可比的CNN和基于注意力的模型.
- 与RetinaNet一起的EMOV2-5M在物体检测方面实现了41.5 mAP,比之前的模型有2.6个改进.
- 通过改进的训练配方,EMOV2-5M达到82.9的Top-1准确度,为5M级模型设置了一个新的基准.
结论:
- EMOv2在各种视觉任务中展示了轻型模型的卓越性能和效率.
- 统一的设计方法有效地弥合了CNN和基于注意力的架构.
- EMOv2推出了500万参数范围内的模型的性能极限,使移动设备的高级功能成为可能.
相关概念视频
Depth Perception and Spatial Vision
913
Depth perception is the ability to perceive objects three-dimensionally. It relies on two types of cues: binocular and monocular. Binocular cues depend on the combination of images from both eyes and how the eyes work together. Since the eyes are in slightly different positions, each eye captures a slightly different image. This disparity between images, known as binocular disparity, helps the brain interpret depth. When the brain compares these images, it determines the distance to an object.
913
Vision
55.3K
Vision is the result of light being detected and transduced into neural signals by the retina of the eye. This information is then further analyzed and interpreted by the brain. First, light enters the front of the eye and is focused by the cornea and lens onto the retina—a thin sheet of neural tissue lining the back of the eye. Because of refraction through the convex lens of the eye, images are projected onto the retina upside-down and reversed.
55.3K
Sight Distance in a Vertical Curve
131
Sight distance on vertical curves is critical in roadway design. It ensures drivers can see far enough ahead to identify and respond to hazards effectively. This directly impacts safety, driver comfort, and the overall efficiency of the transportation network.Vertical curves are classified into crest and sag curves based on their geometry. For crest curves, sight distance is determined by the line of sight between a driver's eye and a small object on the road's surface. Design parameters for...
131
Light Acquisition
8.6K
In order to produce glucose, plants need to capture sufficient light energy. Many modern plants have evolved leaves specialized for light acquisition. Leaves can be only millimeters in width or tens of meters wide, depending on the environment. Due to competition for sunlight, evolution has driven the evolution of increasingly larger leaves and taller plants, to avoid shading by their neighbors with contaminant elaboration of root architecture and mechanisms to transport water and nutrients.
8.6K
Velocity and Position by Graphical Method
8.1K
Velocity and position can be calculated from the known function of acceleration as a function of time. The total area under the acceleration-time graph and the velocity-time graph gives the change in velocity and position, respectively. In the case of an airplane, its acceleration is tracked using the inertial navigation system. The pilot provides the input of the airplane's initial position and velocity before takeoff. The inertial navigation system then uses the acceleration data to...
8.1K
Design Example: Measuring Distance Between Two Points with Obstructions
120
When measuring distances in areas with physical obstructions, such as a lake in a field, surveyors must employ techniques to calculate accurate lengths without direct line measurements. One effective method is the offset technique, which allows for precise distance estimation over inaccessible stretches.In this scenario, a surveyor must measure a side of an area that crosses a lake. Since the measuring tape cannot span the lake, the surveyor begins by establishing a baseline that aligns with...
120


