Most present smartphones use digital camera scene detection to regulate the photograph processing parameters and digital camera settings. Nonetheless, no out there public datasets and fashions have been out there for this job, and every producer has designed its personal restricted answer. Just lately, researchers from ETH Zurich proposed a novel large-scale dataset for scene detection.
It incorporates greater than 11 thousand pictures and 30 scene classes. A number of environment friendly fashions are proposed for the duty. The accuracy of greater than 94% is achieved. An intensive analysis of the proposed answer on smartphones within the wild was carried out. It was confirmed that the advised method works sufficiently quick on trendy units.
The mannequin is performative and strong in varied scenes, circumstances, and environments. The dataset and the designed fashions are actually publicly out there as a way to set up an environment friendly baseline answer for this job.
AI-powered computerized digital camera scene detection mode is these days out there in almost any trendy smartphone, although the issue of correct scene prediction has not but been addressed by the analysis group. This paper for the primary time fastidiously defines this downside and proposes a novel Digital camera Scene Detection Dataset (CamSDD) containing greater than 11K manually crawled pictures belonging to 30 totally different scene classes. We suggest an environment friendly and NPU-friendly CNN mannequin for this job that demonstrates a top-3 accuracy of 99.5% on this dataset and achieves greater than 200 FPS on the current cell SoCs. A further in-the-wild analysis of the obtained answer is carried out to investigate its efficiency and limitation within the real-world eventualities. The dataset and pre-trained fashions used on this paper can be found on the project website.
Analysis paper: Pouget, A., “Quick and Correct Digital camera Scene Detection on Smartphones”, 2021. Hyperlink: https://arxiv.org/abs/2105.07869