How can you choose between cloud-based and edge-based vision services?

Cloud-based vision services refer to online platforms providing diverse functionalities for image and video analysis, including face recognition, object detection, optical character recognition, and scene comprehension. Users can upload visual data to the cloud and access analysis results via an API or web interface. Prominent examples include Google Cloud Vision, Amazon Rekognition, and Microsoft Azure Cognitive Services. These services enable efficient and scalable processing of visual data, facilitating a wide range of applications in various domains.

Cloud-based services have the advantage of providing large storage and processing capacity to the processing. This allows not only to run complex processing on the cloud servers, which would overwhelm an edge computer, but it provides also additional information in form of maps, object shape descriptions, appearances, etc. for recognition tasks. Many do not understand the difference between Image Processing and Computer Vision. Image processing applies operators on the image data to change the resolution, cluster information or enhance specific image properties, like edge detection. This processing does not require external knowledge and it handles large data volumes. Computer Vision adds additional external knowledge to the processing.

Edge-based vision services are solutions running on the edge hardware without external communication. This type of processing has strong limitations in processing power and storage space. It does not depend on the quality of the external communication channel. It is important in industrial applications and in robotics to provide robust service and fail-safe operation under any condition that includes breakdown of the communication. Due to limits on the bandwidth of the communication channel, edge-based vision usually implements image processing algorithms that pre-process the large image data and reduce or group the information that is needed in post-processing. This data is usually sufficient for navigation and obstacle avoidance.

Cloud-based services can provide large storage space for a-priori information and high volume computational capabilities. They are ideal for abstraction of the information that was captured by the camera and often they work on post-processed information that has a smaller volume than original raw camera images. This reduces the data volume that needs to be sent over the network. Additionally, this abstraction services do not need to run with the frame-rate of the camera, which reduces even further the bandwidth requirements on the network. In robotics, cloud-services can only be used for non-essential task to enhance the processing. Low level detects an obstacle and cloud tells what it was.

Cloud-based vision services can become costly, especially as usage scales up, potentially leading to unpredictably high expenses. Storing sensitive visual data in the cloud raises concerns about data protection and compliance with regulations like GDPR or HIPAA.

We can consider the information flow from the camera as a sequence of processing blocks reducing the volume and abstracting the knowledge. Edge computing has the big advantage of running usually directly on the hardware of the camera, which omits the necessity of transferring the large images over the network and it reduces significantly the latency (time between image acquisition and when result is available), which is important in control applications. The processing is robust against network and infrastructure failures.

Edge devices often have limited processing power compared to cloud servers. This limitation can constrain the complexity of algorithms that can be run. They also have limited storage space, which can be a challenge when dealing with large volumes of visual data or when you need to store historical data for analysis.

The big drawbacks are the limited computational power, small storage space and power consumption on especially autonomous agents. Field cameras and mobile robots usually balance the usage of edge- and cloud computing. Essential computing for safety needs usually to run on the edge computer.

A hybrid approach combining both cloud-based and edge-based vision services to leverage the strengths of each while mitigating their respective limitations can sometimes be more fruitful. Here are some of the merits of this approach: - Real-time Efficiency: Edge devices handle time-sensitive tasks - Centralized Analytics and Storage: Cloud-based processing for data aggregation and in-depth analysis - Cost Optimization: Cloud offers scalability and cost efficiency with dynamic resource allocation - Security and Privacy: Sensitive data remains on the edge, enhancing security - Redundancy and Fault Tolerance: Ensures system reliability by shifting tasks in case of edge device failures