Panasonic Holdings Focuses on “overlooked data” in AI Models and Develops New Technology Useful for Deploying AI in Multiple Sites

Osaka, Japan - Panasonic Holdings Corporation (Panasonic Holdings) has developed an AI technology that enables horizontal deployment of AI models with high precision using little training data by focusing on objects that AI models often overlook.

Regarding the on-site implementation of image recognition AI, attention is focused on increasing efficiency through the horizontal deployment of developed models, rather than the individual development of AI models for each site.

However, there is a problem in that recognition performance deteriorates in environments with different image shooting conditions. Generally, improving recognition performance requires a large amount of training data in a new environment, which is costly and time consuming. For this issue, a method called “active domain adaptation” has been proposed, in which AI automatically selects data that it deems likely to be effective for learning, achieving the same effect as labeling all data with only a small amount of real labeling.

The active domain adaptation method preferentially selects data for which the AI model is not confident in its prediction results to efficiently acquire new environmental information. On the other hand, data that is “overlooked” by the AI model is unlikely to be selected as training data, even though it contains important information that prevents data from going undetected.

Therefore, Panasonic Holdings has developed a new algorithm that can take into account data that the AI model has overlooked. In an evaluation experiment*¹ using a benchmark dataset, Panasonic Holdings achieved recognition performance equivalent to labeling all data by labeling just 5% of images in a new environment.

The technology will be presented at the main conference of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2024, as a research outcome of the REAL-AI*² program for developing top human resources in the Panasonic group. The conference will be held in Seattle, Washington, United States, from June 17 to 21, 2024. This conference is one of the top international AI and computer vision conferences.

Panasonic Holdings aims to contribute to helping customers' lives and work through research and development of AI technology that accelerates social implementation and training of top AI researchers.

The active domain adaptation method, which combines two techniques, “active learning” and “domain adaptation,” has been proposed as a method for adapting AI models to environments with different image shooting conditions while reducing cost and time. Active learning selects data with high uncertainty; that is, data for which the AI has no confidence in the predicted results, from a large amount of data taken in a new environment that has not yet been assigned a training data label, as data that the AI wants human operators to attach a training data label to. Since annotation requires a huge amount of cost and time, the cost and accuracy active learning requires greatly depends on how well you select a small number of useful data for annotation. Then, in domain adaptation, the selected data is used as additional training data to tune the AI model. On the other hand, Panasonic Holdings’ preliminary analysis of the causes of accuracy loss in domain adaptation shows that the increase in false negative (FN) errors, or undetected errors, has a significant impact on performance.

In other words, to suppress accuracy degradation in domain adaptation, it is important to design active learning that takes undetected objects into account. The problem with the conventional approach mentioned above is that it is difficult to select data that AI has “confidently” overlooked, leading to undetected data. Therefore, to consider undetected objects, Panasonic Holdings proposed a method that integrates the False Negative Prediction Module (FNPM), which takes into account the possibility that AI will overlook an object (“undetectability”). The introduction of FNPM, which evaluates the undetectability of each image, makes it easier to select images that contain a large number of undetected objects. Since it can increase robustness against undetected data, it is possible to suppress performance degradation in domain adaptation.

In evaluation*¹ using a benchmark data set from in-vehicle cameras, Panasonic Holdings achieved recognition performance equivalent to labeling all data by labeling just 5% of images in a new environment.

This method uses a new technique to focus on data that AI models tend to overlook, and is a technology that allows AI models to be deployed efficiently in multiple sites with fewer labeling steps and associated costs. In particular, we were able to achieve its effectiveness in an evaluation experiment on a benchmark set of in-vehicle cameras, which have large variations in appearance, so we believe that it can be widely applied to object detection tasks. It is expected to contribute to the multi-site deployment of AI models in fields such as on-site optimization solutions in supply chain management, surveillance, and sensing for mobility applications.

*1: Comparison of accuracy when domain adaptation is performed between major benchmark datasets such as BDD100 and KITTI by labeling only a small amount of data that is allowed in a new environment (5%, 1%, etc.)
*2: An in-company group develops top human resources to lead to advanced AI R&D of the Panasonic group through developing top human resources for value creation and rapid business development with advanced technologies, supervised by two professors: Prof. Tadahiro Taniguchi is a professor at Kyoto University and a visiting research professor at Ritsumeikan University; Prof. Takayoshi Yamashita is a professor at Chubu University. Young and senior researchers are challenging top conferences, and many research papers have been accepted at these conferences.

This research is the result of collaboration between Yuzuru Nakamura and Yasunori Ishii of the Panasonic Holdings Technology Division and Takayoshi Yamashita, professor at Chubu University, and was developed within the framework of the Panasonic Group's AI expert training program REAL-AI. At The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR2024), a top conference in the computer vision field where more than 10,000 research papers were submitted, out of 2,719 papers that passed through the narrow gate with an acceptance rate of 23.6%. This research was selected as a "highlight", with only 11.9% (324) of the featured papers selected.

• CVPR2024 Official
  https://cvpr.thecvf.com/Conferences/2024

• Panasonic × AI Website
  https://tech-ai.panasonic.com/en/