Denso ITLAB Presents Three Advanced AI Research Papers at 'CVPR 2026' to Accelerate Mobility Evolution

Denso IT Laboratory, Inc. (ITLAB) (Headquarters: Minato-ku, Tokyo; President and Representative Director: Hirotoshi Iwasaki), an organization dedicated to fundamental R&D for the future of vehicles and mobility, announced that three research papers authored by its researchers have been accepted for the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026 (CVPR 2026), to be held in Denver, Colorado, from June 3 to 7, 2026.

CVPR is one of the world's most authoritative top conferences in computer vision and pattern recognition. This year, 4,090 papers were accepted from 16,092 submissions, an acceptance rate of 25.4%.

The three accepted papers were co-authored with researchers from Denso, Tokyo University of Science, and Kyushu University. All represent research results expected to be applied in mobility sectors such as autonomous driving and Advanced Driver Assistance Systems (ADAS). They were highly evaluated for their originality in rethinking existing method assumptions and proving effectiveness.

Including these three from ITLAB and one from its U.S. subsidiary, DENSO International America, Inc., a total of four papers from the Denso Group will be presented at CVPR 2026.

Overview of Accepted Papers

### Proposal of a New Quantization Method Using Symmetry Hidden in Pre-trained Vision Models

There is a rapidly growing need to run high-precision AI on hardware with limited power and computational resources, from smartphones to in-vehicle systems. 'Quantization' is the technology used to convert a model's internal parameters from high-precision floating-point numbers to integers with fewer bits. While this reduces memory usage and computation, it often lowers recognition accuracy due to quantization errors. Solving this trade-off has been a long-standing challenge.

In response, the joint research team from ITLAB and Denso discovered that by removing a small percentage of outliers (e.g., the top 1%), the remaining distribution of weights in a pre-trained vision model becomes nearly symmetric around zero. This means that by treating outliers separately, the majority of weights can be represented using hardware-efficient symmetric quantization. The team proposed 'DASQ' (Dense and Additive Sparse Quantization) based on this 'hidden symmetry.'

DASQ decomposes weights into 'a majority of symmetric components' and 'a minority of outlier components,' processing them in parallel. This design eliminates the need for zero-point offsets—values that shift the quantization reference point and increase multiplier circuit area by approximately 1.3 times. FPGA evaluation proved that DASQ achieves higher precision and lower power consumption compared to the conventional asymmetric quantization (AsymQ) method.

In autonomous driving and ADAS, high-precision vision models for object recognition and pedestrian detection must operate in real-time on in-vehicle SoCs with strict power constraints. DASQ's design, which integrates algorithms with hardware implementation, is expected to contribute to power-efficient, high-precision AI inference on edge devices including in-vehicle SoCs.

Comments from ITLAB Researchers

### Masafumi Mori (Visiting Researcher, R&D Group, Denso IT Laboratory)

### Shinya Gongyo, Ph.D. (Researcher, R&D Group, Denso IT Laboratory)

At CVPR, the discovery of 'hidden symmetry' in weight distributions, the proposal of the DASQ method, and verification through multiple tasks and FPGA implementation were highly praised. Moving forward, we will accelerate our efforts to enable power-efficient implementation of high-performance vision models on edge devices including in-vehicle AI.