Fast and Accurate Camera Scene Detection on Smartphones

Victoria D. Doty

Most existing smartphones use camera scene detection to adjust the photograph processing parameters and camera configurations. Even so, no offered general public datasets and models ended up offered for this process, and each and every company has designed its own constrained option. Not too long ago, scientists from ETH Zurich proposed a novel significant-scale dataset for scene detection.

Picture credit history: Christian Widell by using Wikimedia

It is made up of much more than eleven thousand illustrations or photos and 30 scene types. Quite a few successful models are proposed for the process. The accuracy of much more than 94{394cb916d3e8c50723a7ff83328825b5c7d74cb046532de54bc18278d633572f} is reached. A extensive analysis of the proposed option on smartphones in the wild was carried out. It was verified that the prompt strategy works adequately speedy on present day equipment.

The product is performative and strong in a variety of scenes, situations, and environments. The dataset and the designed models are now publicly offered in get to create an successful baseline option for this process.

AI-powered automatic camera scene detection mode is today offered in virtually any present day smartphone, nevertheless the trouble of accurate scene prediction has not but been addressed by the research community. This paper for the initial time carefully defines this trouble and proposes a novel Camera Scene Detection Dataset (CamSDD) containing much more than 11K manually crawled illustrations or photos belonging to 30 distinctive scene types. We suggest an successful and NPU-helpful CNN product for this process that demonstrates a top-3 accuracy of 99.five{394cb916d3e8c50723a7ff83328825b5c7d74cb046532de54bc18278d633572f} on this dataset and achieves much more than two hundred FPS on the modern cellular SoCs. An supplemental in-the-wild analysis of the attained option is carried out to analyze its effectiveness and limitation in the real-world scenarios. The dataset and pre-qualified models employed in this paper are offered on the challenge website.

Investigation paper: Pouget, A., “Fast and Precise Camera Scene Detection on Smartphones”, 2021. Connection: muscles/2105.07869

Next Post

Arts & Entertainment Management

This framework consists of the development of visuals that fall outdoors the formal gameplay setting, such as designing the visual branding of your game. Branding will include an iconic title that might be current on the sport’s begin, and can observe through on the coed designed cover art for the […]

Subscribe US Now