Fast and Accurate Camera Scene Detection on Smartphones

Most recent smartphones use digicam scene detection to change the image processing parameters and digicam settings. Yet, no available general public datasets and products ended up available for this task, and every single maker has designed its personal constrained option. Not too long ago, researchers from ETH Zurich proposed a novel substantial-scale dataset for scene detection.

Graphic credit history: Christian Widell by using Wikimedia

It includes extra than 11 thousand pictures and 30 scene categories. Various economical products are proposed for the task. The precision of extra than 94{d11068cee6a5c14bc1230e191cd2ec553067ecb641ed9b4e647acef6cc316fdd} is attained. A comprehensive analysis of the proposed option on smartphones in the wild was carried out. It was confirmed that the prompt technique operates adequately rapid on modern devices.

The model is performative and robust in many scenes, problems, and environments. The dataset and the designed products are now publicly available in buy to build an economical baseline option for this task.

AI-driven automated digicam scene detection manner is at present available in practically any modern smartphone, while the challenge of exact scene prediction has not still been tackled by the analysis neighborhood. This paper for the first time very carefully defines this challenge and proposes a novel Digicam Scene Detection Dataset (CamSDD) that contains extra than 11K manually crawled pictures belonging to 30 various scene categories. We suggest an economical and NPU-helpful CNN model for this task that demonstrates a prime-3 precision of 99.5{d11068cee6a5c14bc1230e191cd2ec553067ecb641ed9b4e647acef6cc316fdd} on this dataset and achieves extra than two hundred FPS on the latest cellular SoCs. An further in-the-wild analysis of the attained option is executed to review its effectiveness and limitation in the serious-entire world scenarios. The dataset and pre-trained products utilized in this paper are available on the task web site.

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