blob: d0dc1e670c23adc232bb120c9727754a356366e3 [file] [log] [blame]
Keir Mierle8ebb0732012-04-30 23:09:08 -07001// Ceres Solver - A fast non-linear least squares minimizer
2// Copyright 2010, 2011, 2012 Google Inc. All rights reserved.
3// http://code.google.com/p/ceres-solver/
4//
5// Redistribution and use in source and binary forms, with or without
6// modification, are permitted provided that the following conditions are met:
7//
8// * Redistributions of source code must retain the above copyright notice,
9// this list of conditions and the following disclaimer.
10// * Redistributions in binary form must reproduce the above copyright notice,
11// this list of conditions and the following disclaimer in the documentation
12// and/or other materials provided with the distribution.
13// * Neither the name of Google Inc. nor the names of its contributors may be
14// used to endorse or promote products derived from this software without
15// specific prior written permission.
16//
17// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20// ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
21// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25// CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26// ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27// POSSIBILITY OF SUCH DAMAGE.
28//
29// Author: David Gallup (dgallup@google.com)
30// Sameer Agarwal (sameeragarwal@google.com)
31
32#include "ceres/canonical_views_clustering.h"
33
Keir Mierle8ebb0732012-04-30 23:09:08 -070034#include "ceres/collections_port.h"
Sameer Agarwal0beab862012-08-13 15:12:01 -070035#include "ceres/graph.h"
Keir Mierle8ebb0732012-04-30 23:09:08 -070036#include "ceres/internal/macros.h"
Sameer Agarwal0beab862012-08-13 15:12:01 -070037#include "ceres/map_util.h"
38#include "glog/logging.h"
Keir Mierle8ebb0732012-04-30 23:09:08 -070039
40namespace ceres {
41namespace internal {
42
43typedef HashMap<int, int> IntMap;
44typedef HashSet<int> IntSet;
45
46class CanonicalViewsClustering {
47 public:
48 CanonicalViewsClustering() {}
49
50 // Compute the canonical views clustering of the vertices of the
51 // graph. centers will contain the vertices that are the identified
52 // as the canonical views/cluster centers, and membership is a map
53 // from vertices to cluster_ids. The i^th cluster center corresponds
54 // to the i^th cluster. It is possible depending on the
55 // configuration of the clustering algorithm that some of the
56 // vertices may not be assigned to any cluster. In this case they
57 // are assigned to a cluster with id = kInvalidClusterId.
58 void ComputeClustering(const Graph<int>& graph,
59 const CanonicalViewsClusteringOptions& options,
60 vector<int>* centers,
61 IntMap* membership);
62
63 private:
64 void FindValidViews(IntSet* valid_views) const;
65 double ComputeClusteringQualityDifference(const int candidate,
66 const vector<int>& centers) const;
67 void UpdateCanonicalViewAssignments(const int canonical_view);
68 void ComputeClusterMembership(const vector<int>& centers,
69 IntMap* membership) const;
70
71 CanonicalViewsClusteringOptions options_;
72 const Graph<int>* graph_;
73 // Maps a view to its representative canonical view (its cluster
74 // center).
75 IntMap view_to_canonical_view_;
76 // Maps a view to its similarity to its current cluster center.
77 HashMap<int, double> view_to_canonical_view_similarity_;
Sameer Agarwal237d6592012-05-30 20:34:49 -070078 CERES_DISALLOW_COPY_AND_ASSIGN(CanonicalViewsClustering);
Keir Mierle8ebb0732012-04-30 23:09:08 -070079};
80
81void ComputeCanonicalViewsClustering(
82 const Graph<int>& graph,
83 const CanonicalViewsClusteringOptions& options,
84 vector<int>* centers,
85 IntMap* membership) {
86 time_t start_time = time(NULL);
87 CanonicalViewsClustering cv;
88 cv.ComputeClustering(graph, options, centers, membership);
89 VLOG(2) << "Canonical views clustering time (secs): "
90 << time(NULL) - start_time;
91}
92
93// Implementation of CanonicalViewsClustering
94void CanonicalViewsClustering::ComputeClustering(
95 const Graph<int>& graph,
96 const CanonicalViewsClusteringOptions& options,
97 vector<int>* centers,
98 IntMap* membership) {
99 options_ = options;
100 CHECK_NOTNULL(centers)->clear();
101 CHECK_NOTNULL(membership)->clear();
102 graph_ = &graph;
103
104 IntSet valid_views;
105 FindValidViews(&valid_views);
106 while (valid_views.size() > 0) {
107 // Find the next best canonical view.
108 double best_difference = -std::numeric_limits<double>::max();
109 int best_view = 0;
110
111 // TODO(sameeragarwal): Make this loop multi-threaded.
112 for (IntSet::const_iterator view = valid_views.begin();
113 view != valid_views.end();
114 ++view) {
115 const double difference =
116 ComputeClusteringQualityDifference(*view, *centers);
117 if (difference > best_difference) {
118 best_difference = difference;
119 best_view = *view;
120 }
121 }
122
123 CHECK_GT(best_difference, -std::numeric_limits<double>::max());
124
125 // Add canonical view if quality improves, or if minimum is not
126 // yet met, otherwise break.
127 if ((best_difference <= 0) &&
128 (centers->size() >= options_.min_views)) {
129 break;
130 }
131
132 centers->push_back(best_view);
133 valid_views.erase(best_view);
134 UpdateCanonicalViewAssignments(best_view);
135 }
136
137 ComputeClusterMembership(*centers, membership);
138}
139
140// Return the set of vertices of the graph which have valid vertex
141// weights.
142void CanonicalViewsClustering::FindValidViews(
143 IntSet* valid_views) const {
144 const IntSet& views = graph_->vertices();
145 for (IntSet::const_iterator view = views.begin();
146 view != views.end();
147 ++view) {
148 if (graph_->VertexWeight(*view) != Graph<int>::InvalidWeight()) {
149 valid_views->insert(*view);
150 }
151 }
152}
153
154// Computes the difference in the quality score if 'candidate' were
155// added to the set of canonical views.
156double CanonicalViewsClustering::ComputeClusteringQualityDifference(
157 const int candidate,
158 const vector<int>& centers) const {
159 // View score.
160 double difference =
161 options_.view_score_weight * graph_->VertexWeight(candidate);
162
163 // Compute how much the quality score changes if the candidate view
164 // was added to the list of canonical views and its nearest
165 // neighbors became members of its cluster.
166 const IntSet& neighbors = graph_->Neighbors(candidate);
167 for (IntSet::const_iterator neighbor = neighbors.begin();
168 neighbor != neighbors.end();
169 ++neighbor) {
170 const double old_similarity =
171 FindWithDefault(view_to_canonical_view_similarity_, *neighbor, 0.0);
172 const double new_similarity = graph_->EdgeWeight(*neighbor, candidate);
173 if (new_similarity > old_similarity) {
174 difference += new_similarity - old_similarity;
175 }
176 }
177
178 // Number of views penalty.
179 difference -= options_.size_penalty_weight;
180
181 // Orthogonality.
182 for (int i = 0; i < centers.size(); ++i) {
183 difference -= options_.similarity_penalty_weight *
184 graph_->EdgeWeight(centers[i], candidate);
185 }
186
187 return difference;
188}
189
190// Reassign views if they're more similar to the new canonical view.
191void CanonicalViewsClustering::UpdateCanonicalViewAssignments(
192 const int canonical_view) {
193 const IntSet& neighbors = graph_->Neighbors(canonical_view);
194 for (IntSet::const_iterator neighbor = neighbors.begin();
195 neighbor != neighbors.end();
196 ++neighbor) {
197 const double old_similarity =
198 FindWithDefault(view_to_canonical_view_similarity_, *neighbor, 0.0);
199 const double new_similarity =
200 graph_->EdgeWeight(*neighbor, canonical_view);
201 if (new_similarity > old_similarity) {
202 view_to_canonical_view_[*neighbor] = canonical_view;
203 view_to_canonical_view_similarity_[*neighbor] = new_similarity;
204 }
205 }
206}
207
208// Assign a cluster id to each view.
209void CanonicalViewsClustering::ComputeClusterMembership(
210 const vector<int>& centers,
211 IntMap* membership) const {
212 CHECK_NOTNULL(membership)->clear();
213
214 // The i^th cluster has cluster id i.
215 IntMap center_to_cluster_id;
216 for (int i = 0; i < centers.size(); ++i) {
217 center_to_cluster_id[centers[i]] = i;
218 }
219
220 static const int kInvalidClusterId = -1;
221
222 const IntSet& views = graph_->vertices();
223 for (IntSet::const_iterator view = views.begin();
224 view != views.end();
225 ++view) {
226 IntMap::const_iterator it =
227 view_to_canonical_view_.find(*view);
228 int cluster_id = kInvalidClusterId;
229 if (it != view_to_canonical_view_.end()) {
230 cluster_id = FindOrDie(center_to_cluster_id, it->second);
231 }
232
233 InsertOrDie(membership, *view, cluster_id);
234 }
235}
236
237} // namespace internal
238} // namespace ceres