With the large-scale implementation of artificial intelligence applications, while the data labeling market is growing rapidly, it is also facing the challenge of labeling costs. According to the IDC report, data labeling accounts for 25% of the time spent in AI application development, and the cost of labeling a piece of data in some medical applications is even as high as 20 yuan . High data accuracy requirements, strong manual dependence, and complex tool usage logic are all core pain points that cause high labeling costs.
For example, an important technology in the field of computer vision -image semantic segmentation can cut pictures into color blocks with different semantics, thereby helping machines understand the entire visual world. In the field of autonomous driving, after the images collected by the on-board camera are processed by the image segmentation algorithm, it can help the car realize functions such as avoiding obstacles and finding drivable areas. Image semantic segmentation technology requires high training data, and the segmentation accuracy usually needs to reach the pixel level. In the real world, the labeling efficiency of irregular object edges will be lower.
▲ Schematic diagram of MatrixGo panoramic semantic segmentation
▲ Appen MatrixGo Platform Toolbox
The intelligence level of the labeling platform
Intelligence Level of Platform
We divide the data labeling platform into five stages from L0 to L4 according to the level of intelligence and automation. Among them, L2 (that is, intelligent interaction) takes into account the interaction of "people" in the data labeling process. Through better algorithm intervention and interactive logic guidance, users can complete efficient annotation through simple operations.
▲ Five intelligent stages from L0 to L4
The annotator first finds the object to be annotated, and after inputting the center point, the model gives the first recognition effect; for the parts that are not identified accurately, the annotator only needs to click the response form to inform the model of the incorrect area predicted. Get corrected recognition results without relying on manual line drawing.
Intelligent interactive process
Process of Intelligent Interaction
By quickly clicking on the foreground object that needs to be marked, the model will predict the outline of the main body; when there is an incorrect mark that needs to be modified, by clicking on the feedback of the background position, the tool will automatically perform intelligent edge closing and erasing operations . During the model recognition process, the annotators can obtain more accurate recognition results through simple interactive input. Relying on simple point selection to replace the dense outline drawing process, according to the actual application statistics of the project, it can save about 50% of the labeling time compared with pure manual labeling.
▲ Interaction flow diagram
By designing a user interaction understanding module , inputting a series of user interaction behaviors, and outputting corresponding high-level features to model the user input behavior patterns, it is possible to effectively infer better quality annotation results. At the same time, it assists in the use of click behavior sampling , and bury points according to the user's historical marking behavior to extract effective data for learning. In terms of output, taking into account different project and business requirements, it supports output in different formats such as general rectangular boxes, polygons, and pixel maps.
▲ Illustration of the implementation method
In 3D recognition scenarios, we also provide interactive intelligent tools for quickly identifying objects in point clouds. The model can quickly identify lane lines within the range delineated by the annotator, and return the final result to the annotator.
▲3D point cloud lane line recognition
Interactive intelligent labeling can greatly improve the efficiency and accuracy of labeling by combining algorithm recognition + manual judgment. Through the full life cycle of AI empowering data acquisition and standardization, the efficiency of data production is improved, and the development of AI applications is fully fed with data, thus providing strong support for the large-scale implementation of AI applications in more scenarios.