Performance Comparison of Bundler and PhotoModel3D

We compare the performance of our PhotoModel3D program with that of Bundler. Bundler is a structure-from-motion system written in C and C++. An earlier version of the Bundler system was used in the Photo Tourism Project of the University of Washington (Dr. Noah Snavely and Professor Steven M. Seitz) and Microsoft (Dr. Richard Szeliski). Bundler is distributed by Professor Noah Snavely under the GNU General Public License. For information on Bundler, please contact Dr. Noah Snavely directly.

Both Bundler and PhotoModel3D are fully-automated, end-to-end 3D modeling systems that take as input digital photographs and produce 3D computer models as output without any user intervention. Input pcitures for the comparison study were taken using hand-held, cosumer-market, point-and-shoot cameras and camera phones. These cameras were not calibrated. No special care was exercised and no photography equipment (e.g., tripod) was used in taking the pictures. These pictures were of different sizes but we down sampled them to a uniform VGA (640x480) size for processing. No other information, other than the VGA-sized pictures, were given to both programs for processing.

Both programs were tested on an Intel Core 2 Duo desktop (2.66GHz, 2GRAM) running Windows Vista (SP2). A single setting of the control script for Bundler and PhotoModel3D was used in all data sets. The control scripts of both Bundler and PhotoModel3D were set up in such a way that only one CPU core was used with no hardware/GPU/DSP acceleration. Bundler results were generated using the v0.3 binary release for Windows (May 4, 2009 release).

While both Bundler and PhotoModel3D infer 3D structure and camera motion information, the real-world data sets we use for comparison do not come with ground-truth camera motion parameters. Our program generates both discrete (a cloud of 3D points) and continuous (textured-mapped surfaces) 3D outputs, while Bundler outputs in the point-cloud format. Hence, we only compare qualitatively the constructed 3D models in the point-cloud format. For each data set, the number of input images, runtime, point-cloud density, and a short movie of the constructed 3D model are provided.




This page demonstrates our core 3D reconstruction technology.
Please visit our demo page for more examples.
Please note that Photomodel3D was developed by VisualSize in-house with no 3rd party licensing or royalty payment required. Please contact us for an in-depth evaluation and customized solutions.


Click on the pictures to see details

For Visualsize's texture-mapped models, please see our demo page

Sample Image

Bundler models

PhotoModel3D models

Statistics



Calculator (5 images) Run time # of 3D points
Bundler 0:21 762
Photomodel3D 0:30 3791
Handprint in dirt (5 images) Run time # of 3D points
Bundler 0:40 4877
Photomodel3D 3:23 23646
Castle (6 images) Run time # of 3D points
Bundler 0:27 244
Photomodel3D 1:07 10974
Castle (7 images) Run time # of 3D points
Bundler 0:31 81
Photomodel3D 0:56 3498
Happy face (9 images) Run time # of 3D points
Bundler 0:40 2041
Photomodel3D 1:39 14660
ET (9 images) Run time # of 3D points
Bundler 0:47 514
Photomodel3D 1:23 8699
Fireplace (10 images) Run time # of 3D points
Bundler 0:51 1422
Photomodel3D 1:18 10644
Wreath (10 images) Run time # of 3D points
Bundler 0:49 1033
Photomodel3D 1:57 13565
Robot (10 images) Run time # of 3D points
Bundler 0:46 172
Photomodel3D 1:02 2262
Computers (10 images) Run time # of 3D points
Bundler 0:45 1755
Photomodel3D 1:26 12355
Kermit (11 images) Run time # of 3D points
Bundler 0:44 623
Photomodel3D 1:09 8649
Bee (11 images) Run time # of 3D points
Bundler 0:33 389
Photomodel3D 0:40 2117
Chair (14 images) Run time # of 3D points
Bundler 0:58 2273
Photomodel3D 2:14 21724
Pigs (15 images) Run time # of 3D points
Bundler 0:55 1301
Photomodel3D 1:16 6543
Knight (16 images) Run time # of 3D points
Bundler 0:59 694
Photomodel3D 1:38 7381
Flowers (18 images) Run time # of 3D points
Bundler 1:48 2816
Photomodel3D 4:16 20949
Bear (19 images) Run time # of 3D points
Bundler 1:16 2360
Photomodel3D 1:34 16340
Vase (23 images) Run time # of 3D points
Bundler 2:43 7344
Photomodel3D 7:06 25586
Fountain (23 images) Run time # of 3D points
Bundler 2:56 7330
Photomodel3D 10:51 50844
Church (24 images) Run time # of 3D points
Bundler 2:29 5851
Photomodel3D 8:54 46689
Xmastree (28 images) Run time # of 3D points
Bundler 2:29 3074
Photomodel3D 6:54 39379
Logcabin (29 images) Run time # of 3D points
Bundler 3:11 2264
Photomodel3D 12:03 40430
Small Vase (31 images) Run time # of 3D points
Bundler 2:40 990
Photomodel3D 5:24 24584
NTU Gate (34 images) Run time # of 3D points
Bundler 4:25 8030
Photomodel3D 9:25 64046
Yarn (42 images) Run time # of 3D points
Bundler 8:43 24583
Photomodel3D 12:03 69548
Dragon (44 images) Run time # of 3D points
Bundler 6:04 8307
Photomodel3D 10:43 25319
Hallmark (50 images) Run time # of 3D points
Bundler 4:19 3023
Photomodel3D 11:47 48737
Tiger (58 images) Run time # of 3D points
Bundler 15:15 2827
Photomodel3D 22:31 77068
More Tiger (64 images) Run time # of 3D points
Bundler 8:48 10247
Photomodel3D 14:55 110543
Shoes (72 images) Run time # of 3D points
Bundler 6:01 5150
Photomodel3D 39:14 56997
Leopard (88 images) Run time # of 3D points
Bundler 18:32 20001
Photomodel3D 33:50 112926



Performance Comparison of Bundler+PMVS2 and PhotoModel3D

It is recommended that Bundler should be used in conjunction with Dr. Yasutaka Furukawa's PMVS2 to generate dense point clouds. Below are some comparison results of PhotoModel3D with Bundler+PMVS2.

Both programs were tested on the same Intel Core 2 Duo Notebook (2.8GHz, 4GRAM) running Windows 7. Bundler used in the comparison study is version 0.4 binary release for Windows (April 10, 2010) and PMVS2 used is version 2 (July 13, 2010).

Again, as no ground-truth camera motion data are available, we only compared qualitatively the constructed 3D models in the point-cloud format. For each data set, the number of input mages, runtime, and short movies of the constructed 3D models for three cases: Bundler only, Bundler + PMVS2, and PhotoModel3D are shown.

Click on the pictures to see details

For Visualsize's texture-mapped models, please see our demo page
Kermit (11 images)
Bundler only Bundler + PMVS2 PhotoModel3D
0:46 1:39 2:03
Bear (19 images)
Bundler only Bundler + PMVS2 PhotoModel3D
1:25 3:18 2:46
Church (24 images)
Bundler only Bundler + PMVS2 PhotoModel3D
3:00 7:24 10:03
Logcabin (29 images)
Bundler only Bundler + PMVS2 PhotoModel3D
3:03 4:10 13:31
Building (16 images)
Bundler only Bundler + PMVS2 PhotoModel3D
1:32 1:56 4:39
Face (22 images)
Bundler only Bundler + PMVS2 PhotoModel3D
1:21 2:05 5:34
Face (21 images)
Bundler only Bundler + PMVS2 PhotoModel3D
1:16 2:16 6:22
Soda Bottle (46 images)
Bundler only Bundler + PMVS2 PhotoModel3D
5:48 11:07 11:18
Bookend (72 images)
Bundler only Bundler + PMVS2 PhotoModel3D
4,177 pts 23,240 pts 187,866