Building Rome in a Day

Sameer Agarwal1, Noah Snavely2 Ian Simon1 Steven M. Seitz1 Richard Szeliski3
1

University of Washington 2 Cornell University 3 Microsoft Research

Abstract
We present a system that can match and reconstruct 3D
scenes from extremely large collections of photographs such
as those found by searching for a given city (e.g., Rome) on
Internet photo sharing sites. Our system uses a collection
of novel parallel distributed matching and reconstruction
algorithms, designed to maximize parallelism at each stage
in the pipeline and minimize serialization bottlenecks. It is
designed to scale gracefully with both the size of the problem
and the amount of available computation. We have experimented with a variety of alternative algorithms at each stage
of the pipeline and report on which ones work best in a
parallel computing environment. Our experimental results
demonstrate that it is now possible to reconstruct cities consisting of 150K images in less than a day on a cluster with
500 compute cores.

1. Introduction
Entering the search term Rome on flickr.com returns more than two million photographs. This collection represents an increasingly complete photographic record of the
city, capturing every popular site, facade, interior, fountain,
sculpture, painting, cafe, and so forth. Most of these photographs are captured from hundreds or thousands of viewpoints and illumination conditionsTrevi Fountain alone
has over 50,000 photographs on Flickr. Exciting progress
has been made on reconstructing individual buildings or
plazas from similar collections [16, 17, 8], showing the potential of applying structure from motion (SfM) algorithms
on unstructured photo collections of up to a few thousand
photographs. This paper presents the first system capable of
city-scale reconstruction from unstructured photo collections.
We present models that are one to two orders of magnitude
larger than the next largest results reported in the literature.
Furthermore, our system enables the reconstruction of data
sets of 150,000 images in less than a day.
To whom correspondence should be addressed. Email: sagarwal@
cs.washington.edu

City-scale 3D reconstruction has been explored previously in the computer vision literature [12, 2, 6, 21] and is
now widely deployed e.g., in Google Earth and Microsofts
Virtual Earth. However, existing large scale structure from
motion systems operate on data that comes from a structured
source, e.g., aerial photographs taken by a survey aircraft
or street side imagery captured by a moving vehicle. These
systems rely on photographs captured using the same calibrated camera(s) at a regular sampling rate and typically
leverage other sensors such as GPS and Inertial Navigation
Units, vastly simplifying the computation.
Images harvested from the web have none of these simplifying characteristics. They are taken from a variety of
different cameras, in varying illumination conditions, have
little to no geographic information associated with them, and
in many cases come with no camera calibration information.
The same variability that makes these photo collections so
hard to work with for the purposes of SfM also makes them
an extremely rich source of information about the world. In
particular, they specifically capture things that people find interesting, i.e., worthy of photographing, and include interiors
and artifacts (sculptures, paintings, etc.) as well as exteriors
[14]. While reconstructions generated from such collections
do not capture a complete covering of scene surfaces, the
coverage improves over time, and can be complemented by
adding aerial or street-side images.
The key design goal of our system is to quickly produce
reconstructions by leveraging massive parallelism. This
choice is motivated by the increasing prevalence of parallel
compute resources both at the CPU level (multi-core) and
the network level (cloud computing). At todays prices, for
example, you can rent 1000 nodes of a cluster for 24 hours
for $10,000 [1].
The cornerstone of our approach is a new system for largescale distributed computer vision problems, which we will
be releasing to the community. Our pipeline draws largely
from the existing state of the art of large scale matching and
SfM algorithms, including SIFT, vocabulary trees, bundle
adjustment, and other known techniques. For each stage
in our pipeline, we consider several alternatives, as some

2. System Design
Our system runs on a cluster of computers (nodes), with
one node designated as the master node. The master node is
responsible for the various job scheduling decisions.
In this section, we describe the detailed design of our
system, which can naturally be broken up into three distinct
phases: (1) pre-processing 2.1, (2) matching 2.2, and (3)
geometric estimation 2.4.

2.1. Preprocessing and feature extraction

We assume that the images are available on a central
store, from which they are distributed to the cluster nodes on
demand in chunks of fixed size. This automatically performs
load balancing, with more powerful nodes receiving more
images to process.
This is the only stage where a central file server required;
the rest of the system operates without using any shared storage. This is done so that we can download the images from
the Internet independent of our matching and reconstruction
experiments. For production use, it would be straightforward
to have each cluster node to crawl the Internet for images at
the start of the matching and reconstruction process.
On each node, we begin by verifying that the image files
are readable, valid images. We then extract the EXIF tags, if
present, and record the focal length. We also downsample
images larger than 2 Mega-pixels, preserving their aspect
ratios and scaling their focal lengths. The images are then
converted to grayscale and SIFT features are extracted from
them [10]. We use the SIFT++ implementation by Andrea
Vedaldi for its speed and flexible interface [20]. At the end of
this stage, the entire set of images is partitioned into disjoint
sets, one for each node. Each node owns the images and
SIFT features associated with its partition.

I0

SIFT

F0

VT
VQ

TF0

VT
VQ

TFK

TFIDF0

...
IK

SIFT

FK

D
F

...
...
IN

SIFT

FN

VT
VQ

TFN

(a)
Round 1

Verify
{F0}

Verify

(b)
Round 2

Verify
{F0}

...

...
MK

...
...

...
...

TFIDFN

MN

(c)

(d)
Merge

M0

TFIDFK

Round 3

Expand1

NM

Distribute k2

The rest of the paper is organized as follows. Section 2

discusses the detailed design choices and implementation of
our system. Section 3 reports the results of our experiments
on three city scale data sets, and Section 4 concludes with a
discussion of some of the lessons learned and directions for
future work.

N0

Distribute k1

algorithms naturally distribute and some do not, and issues

like memory usage and I/O bandwidth become critical. In
cases such as bundle adjustment, where we find that the
existing implementations did not scale, we have created our
own high performance implementations. Designing a truly
scalable system is challenging, and we discovered many
surprises along the way. The main contributions of our
paper, therefore, are the insights and lessons learned, as well
as technical innovations that we invented to improve the
parallelism and throughput of our system.

M0

MN

(e)
Merge

{T0}
...

Verify

Merge

{FN}

{TN}

T0
T0

TC

Merge

CC
{FN}
(f)

(g)

(h)

(i)

(j)

(k)

TC
(l)

(m)

Figure 1. Our multi-stage parallel matching pipeline. The N input

images are distributed onto M processing nodes, after which the
following processing stages are performed: (a) SIFT feature extraction, vocabulary tree vector quantization, and term frequency
counting; (b) document frequency counting; (c) TFIDF computation and information broadcast; (d) computation of TF-based match
likelihoods; (e) aggregation at the master node; (f) round-robin
bin-packing distribution of match verification tasks based on the
top k1 matches per image; (g) match verification with optional
inter-node feature vector exchange; (h) match proposal expansion
based on images found in connected components CC using the
next k2 best matches per image; (i) more distributed verification;
(j) four more rounds of query expansion and verification; (k) track
merging based on local verified matches; (l) track aggregation by
into C connected components; (m) final distribution of tracks by
image connected components and distributed merging.

2.2. Image Matching

The key computational tasks when matching two images
are the photometric matching of interest points and the geometric verification of these matches using a robust estimate
of the fundamental or essential matrix. While exhaustive
matching of all features between two images is prohibitively
expensive, excellent results have been reported with approximate nearest neighbor search. We use the ANN library [3]
for matching SIFT features. For each pair of images, the
features of one image are inserted into a k-d tree, and the
features from the other image are used as queries. For each
query, we consider the two nearest neighbors, and matches
that pass Lowes ratio test are accepted [10]. We use the
priority queue based search method, with an upper bound
on the maximum number of bin visits of 200. In our experience, these parameters offer a good tradeoff between
computational cost and matching accuracy. The matches returned by the approximate nearest neighbor search are then
pruned and verified using a RANSAC-based estimation of
the fundamental or essential matrix [18], depending on the
availability of focal length information from the EXIF tags.

These two operations form the computational kernel of our

matching engine.
Unfortunately, even with a well optimized implementation of the matching procedure described above, it is not
practical to match all pairs of images in our corpus. For a
corpus of 100,000 images, this translates into 5,000,000,000
pairwise comparisons, which with 500 cores operating at 10
image pairs per second per core would require about 11.5
days to match. Furthermore, this does not even take into
account the network transfers required for all cores to have
access to all the SIFT feature data for all images.
Even if we were able to do all these pairwise matches,
it would be a waste of computational effort, since an overwhelming majority of the image pairs do not match. This is
expected from a set of images associated with a broad tag
like the name of a city. Thus, we must be careful in choosing
the image pairs that the system spends its time matching.
Building upon recent work on efficient object retrieval [15, 11, 5, 13], we use a multi-stage matching scheme.
Each stage consists of a proposal and a verification step. In
the proposal step, the system determines a set of image pairs
that it expects to share common scene elements. In the
verification step, detailed feature matching is performed on
these image pairs. The matches obtained in this manner then
inform the next proposal step.
In our system, we use two methods to generate proposals:
vocabulary tree based whole image similarity 2.2.1 and
query expansion 2.2.4. The verification stage is described
in 2.2.2. Figure 1 shows the system diagram for the entire
matching pipeline.
2.2.1

Vocabulary Tree Proposals

Methods inspired by text retrieval have been applied with

great success to the problem of object and image retrieval.
These methods are based on representing an image as a bag
of words, where the words are obtained by quantizing the image features. We use a vocabulary tree-based approach [11],
where a hierarchical k-means tree is used to quantize the
feature descriptors. (See 3 for details on how we build the
vocabulary tree using a small corpus of training images.)
These quantizations are aggregated over all features in an
image to obtain a term frequency (TF) vector for the image,
and a document frequency (DF) vector for the corpus of
images (Figure 1ae). The document frequency vectors are
gathered across nodes into a single vector that is broadcast
across the cluster. Each node normalizes the term frequency
vectors it owns to obtain the TFIDF matrix for that node.
These per-node TFIDF matrices are broadcast across the
network, so that each node can calculate the inner product
between its TFIDF vectors and the rest of the TFIDF vectors.
In effect, this is a distributed product of the matrix of TFIDF
vectors with itself, but each node only calculates the block

of rows corresponding to the set of images it owns. For

each image, the top scoring k1 + k2 images are identified,
where the first k1 images are used in an initial verification
stage, and the additional k2 images are used to enlarge the
connected components (see below).
Our system differs from that of Nister and Stewenius [11],
since their system has a fixed database to which they match
incoming images. They can therefore store the database in
the vocabulary tree itself and evaluate the match score of
an image on the fly. In our case, the query set is the same
as the database, and it is not available when the features
are being encoded. Thus, we must have a separate matrix
multiplication stage to find the best matching images.
2.2.2

Verification and detailed matching

The next step is to verify candidate image matches, and to

then find a detailed set of matches between matching images.
If the images were all located on a single machine, the task of
verifying a proposed matching image pair would be a simple
matter of running through the image pairs, perhaps with
some attention paid to the order in which the verifications
are performed so as to minimize disk I/O. However, in our
case, the images and their feature descriptors are distributed
across the cluster. Thus, asking a node to match the image
pair (i, j) require it to fetch the image features from two
other nodes of the cluster. This is undesirable, as there
is a large difference between network transfer speeds and
local disk transfers. Furthermore, this creates work for three
nodes. Thus, the image pairs should be distributed across the
network in a manner that respects the locality of the data and
minimizes the amount of network transfers (Figure 1fg).
We experimented with a number of approaches with some
surprising results. We initially tried to optimize network
transfers before any verification is done. In this setup, once
the master node has all the image pairs that need to be verified, it builds a graph connecting image pairs which share
an image. Using MeTiS [7], this graph is partitioned into
as many pieces as there are compute nodes. Partitions are
then matched to the compute nodes by solving a linear assignment problem that minimizes the number of network
transfers needed to send the required files to each node.
This algorithm worked well for small problem sizes, but
as the problem size increased, its performance became degraded. Our assumption that detailed matches between all
pairs of images take the same constant amount of time was
wrong: some nodes finished early and were idling for up to
an hour.
The second idea we tried was to over-partition the graph
into small pieces, and to parcel them out to the cluster nodes
on demand. When a node requests another chunk of work,
the piece with the fewest network transfers is assigned to it.
This strategy achieved better load balancing, but as the size

of the problem grew, the graph we needed to partition grew

to be enormous, and partitioning itself became a bottleneck.
The approach that gave the best results was to use a simple
greedy bin-packing algorithm (where each bin represents the
set of jobs sent to a node), which works as follows. The
master node maintains a list of images on each node. When
a node asks for work, it runs through the list of available
image pairs, adding them to the bin if they do not require
any network transfers, until either the bin is full or there are
no more image pairs to add. It then chooses an image (list of
feature vectors) to transfer to the node, selecting the image
that will allow it to add the maximum number of image pairs
to the bin, This process is repeated until the bin is full. This
algorithm has one drawback: it can require multiple sweeps
over all the image pairs needing to be matched. For large
problems, the scheduling of jobs can become a bottleneck.
A simple solution is to only consider a subset of the jobs at a
time, instead of trying to optimize globally. This windowed
approach works very well in practice, and all our experiments
are run with this method.
Verifying an image pair is a two-step procedure, consisting of photometric matching between feature descriptors,
and a robust estimation of the essential or fundamental matrix depending upon the availability of camera calibration
information. In cases where the estimation of the essential matrix succeeds, there is a sufficient angle between the
viewing directions of the two cameras, and the number of
matches are above a threshold, we do a full Euclidean twoview reconstruction and store it. This information is used in
later stages (see 2.4) to reduce the size of the reconstruction
problem.

2.2.3

Merging Connected Components

At this stage, consider a graph on the set of images with

edges connecting two images if matching features were
found between them. We refer to this as the match graph. To
get as comprehensive a reconstruction as possible, we want
the fewest number of connected components in this graph.
To this end, we make further use of the proposals from the
vocabulary tree to try and connect the various connected
components in this graph. For each image, we consider
the next k2 images suggested by the vocabulary tree. From
these, we verify those image pairs which straddle two different connected components (Figure 1hi). We do this only
for images which are in components of size 2 or more. Thus,
images which did not match any of their top k1 proposed
matches are effectively discarded. Again, the resulting image pairs are subject to detailed feature matching. Figure 2
illustrates this. Notice that after the first round, the match
graph has two connected components, which get connected
after the second round of matching.

2.2.4

Query Expansion

After performing two rounds of matching as described above,

we have a match graph which is usually not dense enough to
reliably produce a good reconstruction. To remedy this, we
use another idea from text and document retrieval research
query expansion [5].
In its simplest form, query expansion is done by first
finding the documents that match a users query, and then
using them to query the database again, thus expanding
the initial query. The results returned by the system are
some combination of these two queries. In essence, if we
were to define a graph on the set of documents, with similar
documents connected by an edge, and treat the query as a
document too, then query expansion is equivalent to finding
all vertices that are within two steps of the query vertex.
In our system, we consider the image match graph, where
images i and j are connected if they have a certain minimum
number of features in common. Now, if image i is connected
to image j and image j is connected to image k, we perform
a detailed match to check if image j matches image k. This
process can be repeated a fixed number of times or until the
match graph converges.
A concern when iterating rounds of query expansion is
drift. Results of the secondary queries can quickly diverge
from the original query. This is not a problem in our system,
since query expanded image pairs are subject to detailed
geometric verification before they are connected by an edge
in the match graph.

2.3. Track Generation

The final step of the matching process is to combine all the
pairwise matching information to generate consistent tracks
across images, i.e., to find and label all of the connected
components in the graph of individual feature matches (the
feature graph). Since the matching information is stored locally on the compute node that it was computed on, the track
generation process proceeds in two stages (Figure 1km). In
the first, each node generates tracks from all the matching
data it has available locally. This data is gathered at the
master node and then broadcast over the network to all the
nodes. Observe that the tracks for each connected component of the match graph can be processed independently.
The track generation then proceeds with each compute node
being assigned a connected component for which the tracks
need to be produced. As we merge tracks based on shared
features, inconsistent tracks can be generated, where two
feature points in the same image belong to the same track.
In such cases, we drop the offending points from the track.
Once the tracks have been generated, the next step is to
extract the 2D coordinates of the feature points that occur in
the tracks from the corresponding image feature files. We
also extract the pixel color for each such feature point, which
is later used for rendering the 3D point with the average

33099

33099

35513

33099

35513

129261

118293

92654

92654

2543

2543

41037

2543

41037

100455

100455

60343

109955

64855

109955

64855

66327

66327

39964

Initial Matches

43705
3037
118292
27282
13972

88549

40018

40018
61450

96302

61450

96302

5412

39960
104

116625
48531

86774

39964

86775

48995

Query Expansion 1

48531

86774

39964

86775

48995

96302

5412

39960
104

116625

39964

CC Merge

15277

102389

27282

88549

61450

48531

86774

86775

48995

46262

15277

43705
3037
118292

13972

40018

39960
104

116625

39964

86775

83222

46262

102389

27282

88549

5412

48531

16742

66327

83222

15277

43705
3037
118292

96302

39960
104

116625

66399
101322
88584

86001

16742

46262

13972

61450
5412

122090
117000

66399

102389

40018
96302

86774

3039
122090
117000

83222

27282

88549

61450

48531

86774

25965

3039

66327

43705
3037
118292

13972

40018

39960
104

116625

109955

64855
36587
25965

101322
88584

86001

16742

15277

102389

27282

88549

5412

109955

64855

122090
66399

46262

15277

43705
3037
118292

13972

55720

117000

83222

46262

102389

117980
123361

3039

101322
88584

86001

16742

66327

83222

86005

100455

16743

36586

36587
25965

122090
66399
101322
88584

16742

15275

109954
134888

55760

55720

117000
66399

60343

117980
123361

3039
122090
117000

124710

36586

36587
25965

3039

86001

15275

109954
134888

55760

55720

36587
25965

101322
88584

60343

117980
123361

109955

64855

36587

86001

41359
16743

124710

36586

109954
134888

55720

61722

38055

68091
41359

15275

117980
123361

55760

112136

14757
2990

68091

16743

124710

36586

74555

14757
2990

41359

15275

109954
134888

65125

116995

84355

14757

68091

16743
60343

55760

55720

64871
108233

96301

74555
38055

2990

41359
124710

117980
123361

109956
694

116995

84355

14757

68091

36586

48994

94300
64871

108233

96301

74555
38055

2990

16743
15275

109954
134888

68070
41360
124010
48994

109956
694

116995

41359
60343

55760

100455
68070
41360

94300
64871

108233

96301

84355

14757

68091

57215

124010
48994

109956
694

38055

2990

92505
65119

68070

94300
64871

74555

19115
10492

41360
124010
48994

109956

116995

84355

18990

15276
62039

40017

100455

68070

108233

96301
38055

130620

124011

101945

59314

23977
92505
65119
57215

40017

41360

694

74555

1628
34810
60361
62850

86005

65125
19115
10492

57215

124010

94300
64871

116995

41037

1628

18990

15276
62039
23977
92505

65119

48994

109956

108233

96301

2543

41037
34810

124011

101945

59314

19115
10492

68070

694

84355

8824

60361
62850

86005

65125

14803

41377

73147

49873
10493

1628

18990

15276
62039

40017

41360
124010

94300

124710

2543

41037
34810

124011

101945

59314

23977
92505
65119

4728

91259

8824

60361
62850

86005

65125
19115
10492

57215

40017

18989
22505

43242

73147

49873
10493

1628

18990

15276
62039
23977
92505

65119

87765

61722

91259

34810

124011

101945

59314

65125
19115
10492

57215

112136
18989
22505

43242

8824

60361
62850

86005

18990

15276
62039

70985
61449

87765

61722

73147

49873
10493

1628

124011

101945

40017

693

55060
75015

49824
116997

41377

70985
61449
112136

91259

34810
60361
62850

59314

23977

41361

34812

49030

124013

55060
75015

49824
116997

41377

70985

18989
22505

43242

73147
8824

39962

30036

61720

61448

61449
87765

61722

91259
49873

34299
39962

30036

124013

55060
75015

49824

112136
18989
22505

43242

10493

63807

54469

45356

43870

116997

41377

70985

87765

61722

123345

14470
1231

34299

61448

61449
112136

73147

92654

34321

13942

49030

45356

61720

124013

55060
75015

49824
116997

41377

70985

18989
22505

8824

53569

130620
63807

54469

1231

43870
39962

30036

116844

5433

102896

4728

123345

14470
13942

49030

34299

61448

61449

91259
49873

92654

34321

54469

45356

61720

124013

55060
75015

49824

87765
43242

116843

131056

53569

130620
63807

14470
1231

43870

116997

10493

116844

5433

102896

4728

123345
34321

13942

49030

39962

61448

124013

118294

109346
54467

116843

131056

53569

130620
63807

54469

34299

30036

55061

124191
66328

116844

5433

102896

4728

123345

14470

45356

14803

109346
54467

116843

131056

53569

34321

1231

43870
61720

55764
693

124191
66328

116844

5433

102896

83416
55764

14803

109346
54467

116843

131056

13942

49030

118293
83416

693

124191
66328

109346
54467

130620
63807

39962

30036

55763

118293
55764

14803

124191
66328

4728

123345

54469

34299

61448

22059
34298

83981
34812

83416

693

14803

116844

92654

14470

61720

74763

55763

118293
55764

693

109346

53569

34321

45356

55762

22059
34298

83981
34812

83416
55764

5433

102896

1231

108214
86003
20789

41361

74763

55763

118293
83416

124191

116843

131056

48154
91260

52259

118294

55762

22059
34298

83981
34812

54467

13942

61721
48154

108214
86003
20789

41361

74763

55763

55062

92655

87709
68095

68072

61721

91260

52259

118294

55762

22059
34298

66328

62383

68094
6670
55061

68072

48154

108214
86003
20789

41361

74763

83981
34812

43870

33099

65121

3038

68095

55061

61721

91260

52259

118294

55762

87709

6670

68095

68072

48154

108214
86003
20789

41361

130689

92655
68094

87709

6670
55061

61721

91260

52259

118294

22059

55763

55062

65121

3038

130689

92655
68094

87709
68095

68072

48154

108214

34298

129261
62383

55062

65121

3038

130689

92655
68094
6670
55061

61721

74763

83981

35513

129261
62383

55062

65121

3038

130689

87709
68095

86003

55762

35513

129261
62383

55062

92655
68094
6670

68072

91260

52259

33099

35513

129261
62383

65121

3038

130689

20789

86775

48995

Query Expansion 4

48995

Skeletal Set

Figure 2. The evolution of the match graph as a function of the rounds of matching, and the skeletal set corresponding to it. Notice how the
second round of matching merges the two components into one, and how rapidly the query expansion increases the density of the within
component connections. The last column shows the skeletal set corresponding to the final match graph. The skeletal sets algorithm can
break up connected components found during the match phase if it determines that a reliable reconstruction is not possible, which is what
happens in this case.

color of the feature points associated with it. Again, this

procedure proceeds in two steps. Given the per-component
tracks, each node extracts the feature point coordinates and
the point colors from the SIFT and image files that it owns.
This data is gathered and broadcast over the network, where
it is processed on a per connected component basis.

process. Once a skeletal set is computed, we estimate the

SfM parameters of each resulting component using the incremental algorithm of [16]. The skeletal sets algorithm often
breaks up connected components across weakly-connected
boundaries, resulting in a larger set of components.

2.4. Geometric Estimation

2.4.1

Once the tracks have been generated, the next step is to

run structure from motion (SfM) on every connected component of the match graph to recover a pose for every camera
and a 3D position for every track. Most SfM systems for
unordered photo collections are incremental, starting with a
small reconstruction, then growing a few images at a time,
triangulating new points, and doing one or more rounds
of nonlinear least squares optimization (known as bundle
adjustment [19]) to minimize the reprojection error. This
process is repeated until no more cameras can be added.
However, due to the scale of our collections, running such an
incremental approach on all the photos at once is impractical.
The incremental reconstruction procedure described
above has in it the implicit assumption that all images contribute more or less equally to the coverage and accuracy of
the reconstruction. Internet photo collections, by their very
nature are redundantmany photographs are taken from
nearby viewpoints and processing all of them does not necessarily add to the reconstruction. It is thus preferable to
find and reconstruct a minimal subset of photographs that
capture the essential connectivity of the match graph and the
geometry of the scene [8, 17]. Once this is done, we can
add back in all the remaining images using pose estimation,
triangulate all remaining points, and then do a final bundle
adjustment to refine the SfM estimates.
For finding this minimal set, we use the skeletal sets
algorithm of [17], which computes a spanning set of photographs that preserves important connectivity information
in the image graph (such as large loops). In [17], a two-frame
reconstruction is computed for each pair of matching images
with known focal lengths. In our system, these pairwise
reconstruction are computed as part of the parallel matching

Having reduced the size of the SFM problem down to the

skeletal set, the primary bottleneck in the reconstruction process is the non-linear minimization of the reprojection error,
or bundle adjustment (BA). The best performing BA software available publicly is Sparse Bundle Adjustment (SBA)
by Lourakis & Argyros [9]. The key to its high performance
is the use of the so called Schur complement trick [19] to reduce the size of the linear system (also known as the normal
equations) that needs to be solved in each iteration of the
Levenberg-Marquardt (LM) algorithm. The size of this linear
system depends on the 3D point and the camera parameters,
whereas the size of the Schur complement only depends on
the camera parameters. SBA then uses a dense Cholesky
factorization to factor and solve the resulting reduced linear
system. Since the number of 3D points in a typical SfM
problem is usually two orders of magnitude or more larger
than the number of cameras, this leads to substantial savings. This works for small to moderate sized problems, but
for large problems with thousands of images, computing
the dense Cholesky factorization of the Schur complement
becomes a space and time bottleneck. For large problems
however, the Schur complement itself is quite sparse (a 3D
point is usually visible in only a few cameras) and exploiting
this sparsity can lead to significant time and space savings.
We have developed a new high performance bundle adjustment software that, depending upon the size of the problem,
chooses between a truncated and an exact step LM algorithm.
In the first case, a block diagonal preconditioned conjugate
gradient method is used to solve approximately the normal
equations. In the second case, CHOLMOD [4], a sparse
direct method for computing Cholesky factorization is used
to exactly solve the normal equations via the Schur comple-

Bundle Adjustment

Time (hrs)
Data set
Dubrovnik
Rome
Venice

Images

Cores

Registered

Pairs verified

Pairs found

Matching

Skeletal sets

Reconstruction

57,845
150,000
250,000

352
496
496

11,868
36,658
47,925

2,658,264
8,825,256
35,465,029

498,982
2,712,301
6,119,207

5
13
27

1
1
21.5

16.5
7
16.5

Table 1. Matching and reconstruction statisics for the three data sets.

ment trick. The first algorithm has low time complexity per
iteration, but uses more LM iterations, while the second one
converges faster at the cost of more time and memory per
iteration. The resulting code uses significantly less memory
than SBA and runs up to an order of magnitude faster. The
exact runtime and memory savings depend upon the sparsity
structure of the linear system involved.

2.5. Distributed Computing Engine

Our matching and reconstruction algorithm is implemented as a two-layered system. At the base of the system
is an application-agonstic distributed computing engine. Except for a small core, which is easily ported, the system is
cross-platform and runs on all major flavors of UNIX and
Microsoft Windows. The system is small and extensible,
and while it comes with a number of scheduling policies and
data transfer primitives, the users are free to add their own.
The system is targeted at batch applications that operate
on large amounts of data. It has extensive support for local
caching of application data and on-demand transfer of data
across the network. If a shared filesystem is available, it can
be used, but for maximum performance, all data is stored on
local disk and only transferred across the network as needed.
It supports a variety of scheduling models, from trivially
data parallel tasks to general map-reduce style computation.
The distributed computing engine is written as set of Python
scripts, with each stage of computation implemented as a
combination of Python and C++ code.

3. Experiments
We report here the results of running our system on three
city data sets downloaded from flickr.com: Dubrovnik,
Croatia; Rome; and Venice, Italy. Figures 3(a), 3(b) and 3(c)
show reconstructions of the largest connected components in
these data sets. Due to space considerations, only a sample
of the results is shown here, we encourage the reader to
visit http://grail.cs.washington.edu/rome,
where the complete results are posted, and additional results,
data, and code will be posted over time.
The experiments were run on a cluster of 62 nodes with
dual quad core processors each, on a private network with 1
GB/sec Ethernet interfaces. Each node was equipped with
32 GB of RAM and 1 TB of local hard disk space. The

nodes were running the Microsoft Windows Server 2008

64-bit operating system.
The same vocabulary tree was used in all experiments.
It was trained off-line on 1,918,101 features extracted from
20,000 images of Rome (not used in the experiments reported here). We trained a tree with a branching factor of
10 and a total of 136,091 vertices. For each image, we used
the vocabulary tree to find the top 20 matching images. The
top k1 = 10 matches were used in the first verification stage,
and the next k2 = 10 matches were used in the second component matching stage. Four rounds of query expansion
were done. We found that in all cases, the ratio of number of
matches performed to the number of matches verified starts
dropping off after four rounds. Table 1 summarizes statistics
of the three data sets.
The reconstruction timing numbers in Table 1 bear some
explanation. It is surprising that the SfM time for Dubrovnik
is so much more than for Rome, and is almost the same as
Venice, both which is are much larger data sets. The reason
lies in how the data sets are structured. The Rome and Venice
data sets are essentially a collection of landmarks, which at
large scale have a simple geometry and visibility structure.
The largest connected component in Dubrovnik on the other
hand captures the entire old city. With its narrow alley ways,
complex visibility and widely varying view points, it is a
much more complicated reconstruction problem. This is
reflected in the sizes of the skeletal sets associated with the
largest connected components as shown in Table 2. As we
mentioned above, the skeletal sets algorithm often breaks up
connected components across weakly-connected boundaries,
therefore the size of the connected component returned by
the matching system (CC1), is larger than the connected
component (CC2) that what the skeletal sets algorithm returns.

4. Discussion
At the time of writing this paper, searching on Flickr.com
for the keywords Rome or Roma results in over
2,700,000 images. Our aim is to be able to reconstruct as
much of the city as possible from these photographs in 24
hours. Our current system is about an order of magnitude
away from this goal. We believe that our current approach
can be scaled to problems of this size. However, a number

(a) Dubrovnik: Four different views and associated images from the largest connected component. Note that the component captures the entire
old city, with both street-level and roof-top detail. The reconstruction consists of 4,585 images and 2,662,981 3D points with 11,839,682 observed
features.

Colosseum: 2,097 images, 819,242 points

Trevi Fountain: 1,935 images, 1,055,153 points

Pantheon: 1,032 images, 530,076 points

Hall of Maps: 275 images, 230,182 points
(b) Rome: Four of the largest connected components visualized at canonical viewpoints [14].

San Marco Square: 13,699 images, 4,515,157 points

Grand Canal: 3,272 images, 561,389 points

(c) Venice: The two largest connected components visualized from two different view points each

of open questions remain.

The track generation, skeletal sets, and reconstruction
algorithms are all operating on the level of connected components. This means that the largest few components com-

pletely dominate these stages. Currently, our reconstruction

algorithm only utilizes multi-threading when solving the bundle adjustment problem. While this helps, it is not a scalable
solution to the problem, as depending upon the connectiv-

Data set
Dubrovnik
Rome
Venice

CC1

CC2

Skeletal Set

Reconstructed

6,076
7,518
20,542

4619
2,106
14,079

977
254
1,801

4585
2,097
13,699

Table 2. Reconstruction statistics for the largest connected components in the three data sets. CC1 is the size of the largest connected
component after matching, CC2 is the size of the largest component
after skeletal sets. The last column lists the number of images in
the final reconstruction.

ity pattern of the match graph, this can take an inordinate

amount of time and memory. The key reason we are able
reconstruct these large image collections is because of the
large amount of redundancy present in Internet collections,
which the skeletal sets algorithm exploits. We are currently
exploring ways of parallelizing all three of these steps, with
particular emphasis on the SfM system.
The runtime performance of the matching system depends
critically on how well the verification jobs are distributed
across the network. This is facilitated by the initial distribution of the images across the cluster nodes. An early decision
to store images according to the name of the user and the
Flickr ID of the image meant that most images taken by the
same user ended up on the same cluster node. Looking at the
match graph, it turns out (quite naturally in hindsight) that
a users own photographs have a high probability of matching amongst themselves. The ID of the person who took
the photograph is just one kind of metadata associated with
these images. A more sophisticated strategy would exploit
all the textual tags and geotags associated with the images to
predict what images likely to match and then distribute the
data accordingly.
Finally, our system is designed with batch operation in
mind. A more challenging problem is to start with the output of our system and to add more images as they become
available.

Acknowledgement
This work was supported in part by SPAWAR, NSF grant IIS0811878, the Office of Naval Research, the University of Washington Animation Research Labs, and Microsoft. We also thank
Microsoft Research for generously providing access to their HPC
cluster. The authors would also like to acknowledge discussions
with Steven Gribble, Aaron Kimball, Drew Steedly and David
Nister.

References
[1] Amazon Elastic Computer Cloud. http://aws.amazon.
com/ec2.

[2] M. Antone and S. Teller. Scalable extrinsic calibration of

omni-directional image networks. IJCV, 49(2):143174,
2002.
[3] S. Arya, D. Mount, N. Netanyahu, R. Silverman, and A. Wu.
An optimal algorithm for approximate nearest neighbor
searching fixed dimensions. JACM, 45(6):891923, 1998.
[4] Y. Chen, T. Davis, W. Hager, and S. Rajamanickam. Algorithm 887: CHOLMOD, Supernodal Sparse Cholesky Factorization and Update/Downdate. TOMS, 35(3), 2008.
[5] O. Chum, J. Philbin, J. Sivic, M. Isard, and A. Zisserman.
Total recall: Automatic query expansion with a generative
feature model for object retrieval. In ICCV, pages 18, 2007.
[6] C. Fruh and A. Zakhor. An automated method for large-scale,
ground-based city model acquisition. IJCV, 60(1):524, 2004.
[7] G. Karypis and V. Kumar. A fast and high quality multilevel
scheme for partitioning irregular graphs. SISC, 20(1):359,
1999.
[8] X. Li, C. Wu, C. Zach, S. Lazebnik, and J. Frahm. Modeling
and Recognition of Landmark Image Collections Using Iconic
Scene Graphs. In ECCV, pages 427440, 2008.
[9] M. I. A. Lourakis and A. A. Argyros. Sba: A software
package for generic sparse bundle adjustment. ACM TOMS,
36(1):130, 2009.
[10] D. Lowe. Distinctive image features from scale-invariant
keypoints. IJCV, 60(2):91110, 2004.
[11] D. Nister and H. Stewenius. Scalable recognition with a
vocabulary tree. In CVPR, pages 21612168, 2006.
[12] M. Pollefeys, D. Nister, J. Frahm, A. Akbarzadeh, P. Mordohai, B. Clipp, C. Engels, D. Gallup, S. Kim, P. Merrell, et al.
Detailed real-time urban 3d reconstruction from video. IJCV,
78(2):143167, 2008.
[13] G. Schindler, M. Brown, and R. Szeliski. City-scale location
recognition. In CVPR, pages 17, 2007.
[14] I. Simon, N. Snavely, and S. M. Seitz. Scene summarization
for online image collections. In ICCV, pages 18, 2007.
[15] J. Sivic and A. Zisserman. Video Google: A text retrieval
approach to object matching in videos. In ICCV, pages 1470
1477, 2003.
[16] N. Snavely, S. M. Seitz, and R. Szeliski. Photo Tourism:
Exploring photo collections in 3D. TOG, 25(3):835846,
2006.
[17] N. Snavely, S. M. Seitz, and R. Szeliski. Skeletal graphs for
efficient structure from motion. In CVPR, pages 18, 2008.
[18] P. Torr and A. Zisserman.
Robust computation and
parametrization of multiple view relations. In ICCV, pages
727732, 1998.
[19] B. Triggs, P. McLauchlan, H. R.I, and A. Fitzgibbon. Bundle
adjustment - a modern synthesis. In Vision Algorithms99,
pages 298372, 1999.
[20] A. Vedaldi. Sift++. http://vision.ucla.edu/
vedaldi/code/siftpp/siftpp.html.
[21] L. Zebedin, J. Bauer, K. Karner, and H. Bischof. Fusion
of feature- and area-based information for urban buildings
modeling from aerial imagery. In ECCV, pages IV: 873886,
2008.