The past decade has seen a rise in capturing public spaces for providing well-organized and geo-positioned databases of street view imagery. However, capturing public spaces is challenging, as they contain privacy-sensitive objects, such as faces and license plates. Therefore, these objects must be detected and blurred through an automated process. Although automated methods are labour-free, large resolution images incur high costs for processing. In this research, as we transition from 100 to 250-megapixel system (per cyclorama), we present a framework that reduces the search space of a detection algorithm using depth data obtained from a LIDAR scanner. After then increasing the resolution by 2.5 times and comparing several deep learning architectures, we manage to keep execution time at nearly the same time.
|Title of host publication||IS&T International Symposium on Electronic Imaging|
|Publication status||Published - 2018|
|Event||IS&T International Symposium on Electronic Imaging Science and Technology, : Image Processing: Algorithms and Systems XV - Burlingame, United States|
Duration: 29 Jan 2017 → 2 Feb 2017
|Conference||IS&T International Symposium on Electronic Imaging Science and Technology,|
|Period||29/01/17 → 2/02/17|