From Imagery to Map: Digital Photogrammetric Technologies 15 th International Scientific and Technical Conference From Imagery to Map: Digital Photogrammetric Technologies Photogrammetry and Parallel Computing — Software and Hardware. Andrey Sechin Scientific Director, Racurs Yucatan. Mexico. October 2015
New high productivity sensors Satellite imagery: Higher productivity, large blocks Aerial survey: high overlap, big color depth (>8 bits per channel), special oblique cameras Pleiades: up to sq. km per day VisionMap A3 Edge up to 5000 sq. km per hour (GSD 20 sm)
Computer technologies 64-bit computing Multicore CPU, multithreading Clusters CUDA, MIC,…
Available Hardware CPU cores, Fast Network (10 Gb/s) < $
Orthophoto mosaic algorithm 1.Input the images (Compute pyramid levels and stored in tiles); 2.Orthotransform every image; 3.Compute seamlines; 4.Calculate image statistic for brightness adjustment; 5.Build mosaic divided into sheets (separate files);
Input images What is a tile? Pleiades image (4 tiles)
Orthotransform every image (tile)
Compute seamlines
Calculate image statistic for brightness adjustment
Build mosaic divided into sheets
Some Simple Math Nonparallel part of mosaic algorithm is about 1% T 1 – Single core CPU time α – nonparallel algorithm part T P – Multicore CPU time
Example 1 RESURS-P 100 CPU cores sq. km images area sq. km mosaic area Total time – 50 min Single core time > 2 days
Example 2 UltraCam Eagle 100 CPU cores 510 images/28 sq. km with 3-5 cm GSD Total time – 101 min Single core time > 4 days
Bottlenecks (Storage, Network, RAM) 1.Storage system (random access) ZFS/SAMBA + HDD MB/s, time 120 min OCFS2/iSCSI + 5HDD +RAID100 MB/s, time 30 min OCFS2/iSCSI + single SSD300 MB/s, time 9 min OCFS2/iSCSI + SSD striped600 MB/s (estimated) 2. Network throughput 1 Gbit/s 2 Gbit/s 10 Gbit/s 3. RAM speed DDR -> DDR2 -> DDR3 -> DDR4 ->….. Storage system - mosaic algorithm need sequential access (300 MB/s), some processing stages need random access
Thank you for attention