• Home
  • Blog & News
  • At ICCV2023 presented a Dynamically Controlled Image Signal Processor for Image Recognition (DynamicISP)

2023.10.26

Blog

At ICCV2023 presented a Dynamically Controlled Image Signal Processor for Image Recognition (DynamicISP)

We presented the DynamicISP at ICCV 2023, a top conference in the field of computer vision. This technique allows image recognition AI to control the camera parameters so that AI can recognize images easily, thereby enabling accurate image recognition under severe environmental conditions such as dark places, HDR, and camera shake.

The paper submitted by the R&D team of Technology Infrastructure Center, Sony Group Corporation, was accepted at ICCV 2023, a top conference in the field of computer vision. We presented the paper in Paris from October 2 to 6. ICCV 2023 comprised not only the main conference for paper presentations but also various other programs including tutorials, workshops, and corporate exhibitions. It was a successful event with a turnout of over 7000 people.

Sony Group’s exhibition booth

Opening session

The technology presented at ICCV 2023 allows the image recognition AI to control the camera parameters dynamically so that AI can recognize images easily. Images in formats like PNG and JPG that you often see are created by the Image Signal Processor (ISP) performing various kinds of processing on the values captured by the image sensor. Usually, the ISP is tuned by experts to produce images that appear natural and beautiful to the human eye.

In contrast, our research is based on the following two hypotheses:
1. Images that appear natural to the human eye are different from images that are easily recognizable by AI.
2. The ISP parameters should be changed dynamically according to environmental conditions, rather than being statically tuned.
AI extracts data concerning how the image recognizer in the later stage feels about an image and controls the ISP parameters based on the extracted data in order to create an easily recognizable image.

This has improved the recognition accuracy under severe environmental conditions such as dark places, HDR, and camera shake.

For details of this technique, see below.

At the poster session of the conference, we had a chance to talk with a lot of people, which was a very valuable experience.

Discussing various cutting-edge technologies with excellent experts helps to promote technological advances and idea generation and to increase motivation. It is one of the huge advantages of participating in this conference.


Lastly, the authors of this paper are mentioned below. Thank you very much for reading.

Masakazu Yoshimura, Junji Otsuka, Atsushi Irie, Takeshi Ohashi (from left to right)

Share