Panasonic HD announces two papers accepted at CVPR 2026

Osaka, Japan, May 28, 2026 – Panasonic Holdings Corporation (Panasonic HD) today announced that two of its research papers have been accepted for presentation at CVPR (IEEE/CVF Conference on Computer Vision and Pattern Recognition) 2026, one of the world’s leading international conferences in AI and computer vision. One of the two papers was also selected as a “Highlight” in recognition of its outstanding quality. The company will present the papers at the conference scheduled to be held in Colorado, United States, from June 3 to June 7, 2026.

＜Overview of Accepted Papers＞

■High-Efficiency Spatial Recognition Technology Supporting the Practical Use of Physical AI

This paper presents a technology for efficiently compressing 3D spatial information, enabling both reduced information processing requirements and high spatial recognition capability. The technology is expected to contribute to advances in robotics and physical AI, where AI systems operate in the real world.

Background

In recent years, growing attention has been paid to physical AI, which enables robots and machines to recognize real-world environments, make decisions, and act autonomously. Achieving this requires advanced spatial recognition capabilities, such as understanding positional relationships among objects, and further progress in multimodal AI*¹ is therefore highly anticipated. However, conventional multimodal AI-based spatial recognition technologies tend to require increasing computational cost in order to retain spatial information.

Key Features

Figure 1. Overview of the high-efficiency spatial recognition technology “Proxy3D”.

By combining highly efficient compression of feature representations through clustering with staged spatial recognition learning, this technology reduces the amount of spatial information handled by multimodal AI while delivering spatial recognition performance that is equal to or better than other methods*². For example, while some conventional 3D spatial recognition methods input approximately 8,000 tokens of spatial information into multimodal AI, this technology represents 3D space using 700 tokens. It is expected to support practical applications across a wide range of fields requiring 3D spatial recognition and understanding of positional relationships, as well as future real-time processing by AI systems operating in the physical world.

Paper Information

Title: Proxy3D: Efficient 3D Representations for Vision-Language Models Via Semantic Clustering and Alignment
arXiv: https://arxiv.org/abs/2605.08064
CVPR 2026: https://cvpr.thecvf.com/virtual/2026/poster/39196
This research was developed under the BAIR Open Research Commons*³ initiative led by UC Berkeley and reflects collaboration among researchers from Tsinghua University, Panasonic R&D Company of America, Panasonic HD, and the University of California, Berkeley

■Portable Active Learning (PAL), Reducing the Time and Cost Required for AI Development

This paper presents a technology that achieves high object detection performance with less labeling work. At CVPR 2026, it was selected as a Highlight in recognition of its originality, technical maturity, and future potential.

Background

AI-based image recognition has rapidly expanded across applications such as autonomous driving, factory inspection, and surveillance systems. However, developing high-performance AI models requires large amounts of manually labeled image data (annotation), where people must identify and mark objects in images — a process that is both time-consuming and costly.

Key Features

To address this challenge, Panasonic HD developed “Portable Active Learning (PAL),” a technology that automatically identifies which images are most valuable for AI training. This allows AI models to improve their accuracy while requiring significantly less manual labeling effort.
A key advantage of PAL is that it can be applied to a wide variety of object detection AI models without modifying their internal structure. Across multiple datasets and AI models, PAL achieved the same or better detection accuracy while reducing labeling requirements, with the prior state-of-the-art method requiring approximately 20% more annotation on average*⁴. This technology can help accelerate cost-effective AI deployment in areas such as autonomous driving, edge AI, infrastructure inspection, and factory automation.

Figure 2. Overview of the data selection process using PAL.

Paper Information

Title: Portable Active Learning for Object Detection
arXiv: https://arxiv.org/abs/2605.10349
CVPR 2026: https://cvpr.thecvf.com/virtual/2026/poster/38968
This research was conducted at Panasonic R&D Center Singapore, a research center of Panasonic HD.

Panasonic HD will continue to accelerate the implementation of AI in society and promote research and development of AI technologies that will contribute to improving our customers' lives and workplaces.

Notes:

*1 AI that can process multiple types of information, including images and text, at the same time.

*2 On the spatial reasoning benchmark VSI-Bench, the proposed method achieved an average score of 47.0, representing an improvement of 14.0 points (approximately 42% relative increase) over a comparable multimodal AI model (Qwen2.5-VL-7B). Among the open-source models evaluated in the paper, it demonstrated the second-highest recognition performance, following Spatial-MLLM (48.4).

*3 An AI research institute established as an open collaboration platform bringing together world-leading researchers across industry and academia. As of May 2026, it includes Panasonic Holdings, along with 16 participating companies such as Google and Meta.

*4 PAL was evaluated on COCO, PASCAL VOC, and BDD100K using object detectors including RetinaNet, Faster R-CNN, SSD, YOLOX-Tiny, and YOLO11s, achieving comparable or better performance than prior methods. The prior state-of-the-art PPAL required 20.7% more annotations on average than PAL on COCO and PASCAL VOC using RetinaNet.

Related Information:

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

- Panasonic R&D Center Singapore
https://research.sg.panasonic.com/

About the Panasonic Group

Founded in 1918, and today a global leader in developing innovative technologies and solutions for wide-ranging applications in the consumer electronics, housing, devices, B2B solutions and energy sectors worldwide, the Panasonic Group switched to an operating company system on April 1, 2022 with Panasonic Holdings Corporation serving as a holding company. The Group reported consolidated net sales of 8,458.2 billion yen for the year ended March 31, 2025. To learn more about the Panasonic Group, please visit: https://holdings.panasonic/global/

The content in this website is accurate at the time of publication but may be subject to change without notice.
Please note therefore that these documents may not always contain the most up-to-date information.
Please note that German, Spanish and Chinese versions are machine translations, so the quality and accuracy may vary.