What is the main feature of the new object detection correction technology developed by Hitachi?

It can correct detection results post-hoc by analyzing overall and region-specific features together, without requiring retraining or modification of the existing object detection AI.

How much accuracy improvement can be expected by adopting this technology?

Verification on multiple public benchmarks shows up to a 50% or more improvement in detection accuracy for latest models like Grounding DINO and LLMDet.

What is the additional processing time required for the correction?

It is approximately 0.1 seconds per image in a standard PC environment equipped with an Intel Core i9 CPU and RTX 2080 Ti GPU.

What types of object detection AI can this technology be applied to?

Since it is model-agnostic, it can be applied to standard open-source models as well as black-box AIs accessed via APIs, such as generative AI services.

When and where will this research result be presented?

It is scheduled to be presented in the Findings Track of CVPR 2026, held from June 3 to 7, 2026, under the paper title 'DetRefiner'.

Hitachi Develops DetRefiner to Correct Object Detection AI Results Without Retraining

Published Jun 5, 2026 11:00 AM ・ Updated Jun 14, 2026 12:07 AM ・ 2 min read ・ Source: PR TIMES

Hitachi developed DetRefiner to correct object detection AI results without retraining. By integrating image-wide and region features, it boosts accuracy by up to 50% with only 0.1s overhead, and supports APIs.

Hitachi, Ltd. has developed DetRefiner (Model-Agnostic Detection Refinement with Feature Fusion Transformer), a new AI technology that corrects results from existing object detection models without retraining or modification. By integrating overall image features and local region features using a feature fusion module, DetRefiner corrects detection errors, improving accuracy by up to 50% on benchmarks like COCO, LVIS, ODinW13, and Pascal VOC for models such as Grounding DINO and LLMDet. The system runs efficiently, adding only about 0.1 seconds of processing time per image under a standard PC environment (Intel Core i9 and NVIDIA GeForce RTX 2080 Ti). DetRefiner operates independently of the AI model's internal structure and weights, making it applicable to black-box APIs. Hitachi plans to deploy it as a core technology for Lumada 3.0 to enhance image recognition in manufacturing, facility maintenance, infrastructure monitoring, and aerial analysis. The research paper, authored by Soichiro Okazaki, Tatsuya Sasaki, and Hiroki Ohashi, will be presented at the Findings Track of the CVPR 2026 international conference from June 3 to 7, 2026.

FACT BOX

Source: PR TIMES
Category: business

Hitachi Develops DetRefiner to Correct Object Detection AI Results Without Retraining

⚡ Key Points

FACT BOX

Editorial & Verification Standards

FAQ

Cite this article — HOW TO CITE

AI CRAWLER ACTIVITY