Ronin Test Suite stitched by Autopano Giga Alpha 1.9
This 10x10 Benchmark image suite was stiched by the Alpha 1 release of AutopanoGiga by Autopano.net. I could find no stitchng erros. The watermark is because the program is in 'Trial' mode. And try it I did. Benchmarked 8 different computer configurations and compared them with the Gigapan Stitcher. Having an older computer system I was having enormous difficulty actually getting through any large stitching projects, so I recently build a new state of the art (for this month) computer system and it is a speed demon compared to my previous hardware. Having seen very significant improvements in my stitching speed, I wanted to know which elements in the new system gives rise to the new stellar performance as I did not know if it is processor, memory, or drive speed related. I know that many of the Gigapan users would like to optimise their computer systems for stitching, so I worked on a thorough study of the variables I could manipulate. I was particularly interested in the following in relation to total stitching time: System Memory Size, options tested were 2GB, 4GB, and 8GB System Memory Speed, in the 4GB configuration, tested between running the two sticks of memory at 800 and 1033 MHz memory clock speeds. Hard Drive Type, tested between stitching (with source images, destination image, and system cache all on the target drive) between a single 300GB Seagate 7200 RPM SATA drive and a RAID 0 (Stripping) array consisting of four Western Digital Velociraptor 10,000 RPM 150GB drives Stitching Software, tested Gigapan Stitcher 0.4.3510 and Autopano Giga Alpha Release 1.90 The image set was Randy's Ronin Sculpture set, a 10x10 image array that produces a .5 gigapixel image. Method: My base system consists of a Intel Quad Core q9550 running at stock 2.83GHz set into a MSI P45 Platinum motherboard (Intel Chipset) and Geforce 8800 GTS GPU, 850 watt power supply and Corsair PC8500 4x2GB memory with Windows Vista Ultimate as OS. During all tests I had the Windows Vista Task Manager and Performance Monitor running and displayed. When testing between the different hard drives, I had the source images and destination image or directory on the same drive and assigned Windows to use that drive for Virtual Memory Swap Space. As it was a single stick of memory, when running in 2GB the system was not running Dual Channel. When running in 4GB and 8GB the memory was populated for Dual Channel. Timing of all stitches was done by looking at the file creation and modified time for the created Autopano .PSB images and calculating the difference. For the Gigapan stitches I looked at the 'Additional Info' tab. Only one sample of each configuration was performed. Minimal other tasks were occurring on the system during the tests. I pulled out and plugged in memory modules between the sets and rebooted the computer for each change in memory configuration and drive swap space location change. Here are the results. Autopano 2Gb memory @ 800 MHz RAID drive: 717 seconds Autopano 2Gb memory @ 800 MHz SINGLE drive: 1315 seconds Autopano 4Gb memory @ 800 MHz RAID drive: 707 seconds Autopano 4Gb memory @ 800 MHz SINGLE drive: 1278 seconds Autopano 4Gb memory @ 1066 MHz RAID drive: 702 seconds Autopano 4Gb memory @ 1066 MHz SINGLE drive: 1335 seconds Autopano 8Gb memory @ 800 MHz RAID drive: 718 seconds Autopano 8Gb memory @ 800 MHz SINGLE drive: 1225 seconds Gigapan 2Gb memory @ 800 MHz RAID drive: 2592 seconds Gigapan 2Gb memory @ 800 MHz SINGLE drive: 2890 seconds Gigapan 4Gb memory @ 800 MHz RAID drive: 2430 seconds Gigapan 4Gb memory @ 800 MHz SINGLE drive: 2972 seconds Gigapan 4Gb memory @ 1066 MHz RAID drive: 2411 seconds Gigapan 4Gb memory @ 1066 MHz SINGLE drive: 3025 seconds Gigapan 8Gb memory @ 800 MHz RAID drive: 2406 seconds Gigapan 8Gb memory @ 800 MHz SINGLE drive: 2644 seconds Autopano set to use just one core, 8Gb memory @ 800 MHz RAID drive: 2480 seconds (statistically the same as the Gigapan Stitcher at same machine specs and just a little less than four times longer then when run with four cores) My old system with 1 CPU AMD 3200+ 64 Bit Windows Ultimate and 500Gb 7200 RPM Western Digital Drive, 2GB DDR2 200 memory using the Gigapan stitcher: 7400 seconds With the results statistically analyzed with JMP software: click this link to view full image http://farm4.static.flickr.com/3053/2931692487_43ed0831c2_o.jpg Discussion: Most important factor of all was the number of cores running (and Autopano is the only multicore enabled mosaic stitcher at the present time). The Autopano maxes out all of the CPU's cores, the Gigapan only uses one. The RAID drive also has a very significant effect, especially with the Autopano software (which, using multiple cores gets limited more by the disk subsystem than the processor). It does not provide as grand of improvements with the Gigapan stitcher probably because the system when running Gigapan Stitcher is being limited more by the processor rather than the disk subsystem. For Autopano the RAID array provided a 80-100% boost in speed but the Gigapan stitcher only gets about a 25% boost. I have not tried the Autopano stitcher with the single core option and the single drive but imagine it would have the same performance as the Gigapan Stitcher with the same setup. The two programs seem to produce the same result. I did not ever find any glaring errors in the stitched output. There does not seem to be any big differences in efficiency between the two program as per stitching. Of course the Gigapan stitcher lets you upload right from the stitcher. The Autopano would require you to fire up the Gigapan uploader. But on the bright side, you don't have to take the time to export your panoramas (which I always do) as they are already saved as .PSD or .PSB files. I am a big fan of the .PSB format, and Autopano can include the projected images as separate placed layers too, best for getting rid of those ghosts (although PTGui does a better job with providing editable masking in its .PSB exports). With the Autopano stitcher you can crop and change projection, etc. before rendering, so you can avoid a time consuming task of loading up the image in Photoshop to do so post rendering. The amount of memory was not a significant player between any of the tests, I guess with a 100 image stitch, 2Gb memory is enough (I did not try 1Gb and did not have 16Gb to try with, maybe if it could keep the whole process and images in memory it could fit in 16Gb, but I bet that a large 1000 image pano would overwhelm even that amount of memory pretty fast.) The memory speed had no effect when running at 4Gb memory. No significant improvement was found between any of the tests when running at a memory clock of 800 and 1066 (so what is the point of the faster memory I wonder?) I find it curious how incredibly much slower my 'old' system was, as even though it was a single core AMD, it was running a higher clock frequency than my new multicore processor, so when using a single core program and the same disk system, I would have though they would have been more comparable. Its probably to do with the size of the onchip cache between the 5 year old and the new processor. I should try sticking the old 200 speed memory in and see what stitching time it has. Conclusion: Get the fastest processor you can get and pair it up with a RAID array. Use multicore stitching when and if you can (the Autopano Giga is still in Alpha testing stage and has significant bugs, none of which did I encounter, but their forum is full of them) Hopefully the Gigapan stitcher will soon be multicore ready????!!!! I hope this helps you! I dont ever want to stitch the same image set 18 times again;)
Uploaded Oct 11, 2008 by Jason Buchheim username:odyssey copyright 2008
29,835px by 21,757px, 0.65 Gigapixels

	Unmounted					Acrylic Museum Mount
Purchase Prints	Matte	Glossy	Lustre	Metalic	Deep Glossy	Deep Glossy
12" x 16"
18" x 25"
24" x 33"
30" x 41"
36" x 49"

VIEW OPTIONS please select the approximate angle of view of the image
Flat   1° 10° 45° 90° 170°
Sphere 45° 90° 180° 270° 360°


Odyssey Expeditions Panoram.io VR billboard style