diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index b9cae62..c9cf70f 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -145,5 +145,6 @@ ELSE()
   ADD_EXECUTABLE(stereo_slam stereo_slam.cpp)
   ADD_LIBRARY(${PROJECT_NAME} SHARED ${SOURCES})
 ENDIF()
+ADD_EXECUTABLE(create_dictionary create_dictionary.cpp)
 
 INCLUDE_DIRECTORIES(${INCLUDE_DIRS} ${PCL_INCLUDE_DIRS})
diff --git a/src/create_dictionary.cpp b/src/create_dictionary.cpp
new file mode 100644
index 0000000..cc3e04e
--- /dev/null
+++ b/src/create_dictionary.cpp
@@ -0,0 +1,255 @@
+// This file is part of ScaViSLAM.
+//
+// Copyright 2011 Hauke Strasdat (Imperial College London)
+//
+// ScaViSLAM is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Lesser General Public License as published
+// by the Free Software Foundation, either version 3 of the License, or
+// any later version.
+//
+// ScaViSLAM is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU Lesser General Public License for more details.
+//
+// You should have received a copy of the GNU Lesser General Public License
+// along with ScaViSLAM.  If not, see <http://www.gnu.org/licenses/>.
+
+#include <iostream>
+#include <tr1/unordered_set>
+#include <tr1/unordered_map>
+
+#include <boost/filesystem.hpp>
+#include <boost/regex.hpp>
+#include <boost/dynamic_bitset.hpp>
+
+#include <opencv2/opencv.hpp>
+#include <opencv2/flann/flann.hpp>
+
+
+// Thanks a lot to Adrien Angeli for all help and discussion concerning
+// place recognition using "bag of words".
+
+using namespace std;
+
+list<string> preprocessFiles(const boost::filesystem::path & directory)
+{
+  list<string> name_list;
+
+  if( exists( directory ) )
+  {
+    boost::filesystem::directory_iterator end ;
+    for(boost::filesystem::directory_iterator iter(directory);
+        iter != end ; ++iter )
+    {
+      if (is_directory( *iter)==false )
+      {
+        boost::filesystem3::path  name = iter->path().filename();
+        name_list.push_back(name.string());
+      }
+    }
+  }
+  return name_list;
+}
+
+cv::Mat
+loadImage(const string & img_name)
+{
+  int MAX_WIDTH = 640;
+  int MAX_HEIGHT = 480;
+  int MAX_AREA = MAX_WIDTH*MAX_HEIGHT;
+
+  cout << "load image: " << img_name << endl;
+  cv::Mat img = cv::imread(img_name, 1);
+
+  while(img.size().area()>MAX_AREA)
+  {
+    cv::pyrDown(img, img);
+    cout << "Downsample!" << endl;
+  }
+  cout << "Final size: " << img.size().width <<" "<< img.size().height << endl;
+  cv::Mat img_mono;
+  cv::cvtColor(img, img_mono, cv::COLOR_BGR2GRAY);
+  cout << endl;
+  return img_mono;
+}
+
+bool
+computeKeypoints(const cv::Mat & img_mono,
+                 vector<cv::KeyPoint>  * key_points)
+{
+  double thr = 300;
+  bool failed = false;
+  int trials = 0;
+
+  cout << "Detect SURF features:" << endl;
+
+  while(true)
+  {
+    cv::SurfFeatureDetector surf(thr, 2);
+    surf.detect(img_mono, *key_points);
+    int num_keypoints = key_points->size();
+
+    cout << "Trial: " << trials << "; threshold: " << thr;
+    cout << "; no features: " << num_keypoints << endl;
+
+    if (num_keypoints>2000)
+    {
+      if (num_keypoints>10000)
+      {
+        failed = true;
+        break;
+      }
+      thr += 200;
+    }
+    else if (num_keypoints<500)
+    {
+      thr -=50;
+    }
+    else
+    {
+      break;
+    }
+    if (trials>=5)
+    {
+      failed = true;
+      break;
+    }
+    ++trials;
+
+    if (failed)
+    {
+      return false;
+    }
+  }
+  cout << endl;
+  return true;
+}
+
+void
+computeDescriptors(const cv::Mat & img_mono,
+                   vector<cv::KeyPoint> * key_points,
+                   cv::Mat * descriptors)
+{
+  cv::Mat desc;
+  cv::SurfDescriptorExtractor surf_extr(2);
+  surf_extr.compute(img_mono, *key_points, desc);
+  for (int row=0; row<desc.rows; ++row)
+  {
+    descriptors->push_back(desc.row(row));
+  }
+}
+
+void
+calculateWordsAndSaveThem(int TARGET_NUM_WORDS,
+                          const cv::Mat & descriptors)
+{
+  cout << "Creating up to " << TARGET_NUM_WORDS << " clusters/words..." << endl;
+  cout << "... " << endl;
+  cvflann::KMeansIndexParams kmeans(32, 11, cv::flann::FLANN_CENTERS_KMEANSPP);
+  cv::Mat centers(TARGET_NUM_WORDS, descriptors.cols, CV_32F);
+  typedef cv::flann::L2<float> distance;
+  typedef distance::ResultType DistanceType;
+  typedef distance::ElementType ElementType;
+  cvflann::Matrix<ElementType>
+      flann_features((ElementType*)descriptors.ptr<ElementType>(0),
+                     descriptors.rows, descriptors.cols);
+  cvflann::Matrix<DistanceType>
+      flann_centers((DistanceType*)centers.ptr<DistanceType>(0),
+                    centers.rows, centers.cols);
+  int num_centers
+      = ::cvflann::hierarchicalClustering<distance>(flann_features,
+                                                    flann_centers,
+                                                    kmeans,
+                                                    distance());
+  cout << "Done: dictionary of " << num_centers << " words created!" << endl;
+  assert(sizeof(float)==4);
+  cv::Mat centers_floats_as_4uchars(num_centers,
+                                    descriptors.cols*4,
+                                    CV_8U,
+                                    centers.data);
+  stringstream str_stream;
+  str_stream<< "surfwords" << num_centers << ".png";
+  cv::imwrite(str_stream.str(), centers_floats_as_4uchars);
+
+  cout << "Saved as file: " << str_stream.str() << endl;
+}
+
+void
+createDictionary(const string & base_str,
+                 int MAX_NUM_IMAGES,
+                 int TARGET_NUM_WORDS)
+{
+  list<string> name_list
+      = preprocessFiles(base_str);
+
+  int num_processed_images = 0;
+  cv::Mat descriptors;
+  for (list<string>::iterator it = name_list.begin(); it!=name_list.end(); ++it)
+  {
+    stringstream sst;
+    sst << base_str << *it;
+    cv::Mat img_mono = loadImage(sst.str());
+    vector<cv::KeyPoint> key_points;
+
+    bool success = computeKeypoints(img_mono, &key_points);
+
+    if (success==false)
+    {
+      cout << "abort!" << endl;
+      cout << endl;
+      continue;
+    }
+    computeDescriptors(img_mono, &key_points, &descriptors);
+
+    cout << "Image processed: " << num_processed_images;
+    cout << " of max. " << MAX_NUM_IMAGES << endl;
+    cout << "Number of features: " << descriptors.rows;
+    cout << " (TARGET_NUM_WORDS: " << TARGET_NUM_WORDS << ")" << endl;
+    cout << endl;
+    cout << endl;
+
+    ++num_processed_images;
+    if(num_processed_images>MAX_NUM_IMAGES)
+      break;
+  }
+
+  if (descriptors.rows<TARGET_NUM_WORDS*10)
+  {
+    cout << "ERROR: By far not enough features detected to calculate "
+         << TARGET_NUM_WORDS << " words/clusters!" << endl;
+    exit(0);
+  }
+  calculateWordsAndSaveThem(TARGET_NUM_WORDS,
+                            descriptors);
+}
+
+int main(int argc, const char* argv[])
+{
+  if (argc<2)
+  {
+    cout << "USAGE: create_dictionary FOLDER_WITH_IMAGES "
+         << "[MAX_NUM_IMAGES] [TARGET_NUM_WORDS]"
+         << endl;
+    cout << endl;
+    exit(0);
+  }
+
+  string base_str(argv[1]);
+
+
+  int MAX_NUM_IMAGES = 150;
+  if (argc>=3)
+    MAX_NUM_IMAGES = atoi(argv[2]);
+  int TARGET_NUM_WORDS = max(1000, MAX_NUM_IMAGES*10);
+  if (argc>=4)
+    TARGET_NUM_WORDS = atoi(argv[3]);
+
+  cout << endl;
+  cout << "MAX_NUM_IMAGES: " << MAX_NUM_IMAGES << endl;
+  cout << "TARGET_NUM_WORDS: " << TARGET_NUM_WORDS << endl;
+  cout << endl;
+
+  createDictionary(base_str, MAX_NUM_IMAGES, TARGET_NUM_WORDS);
+}
+
diff --git a/src/dense_tracking.cpp b/src/dense_tracking.cpp
index 0a1bdb6..cd3f32c 100644
--- a/src/dense_tracking.cpp
+++ b/src/dense_tracking.cpp
@@ -39,10 +39,18 @@ ::DenseTracker(const FrameData<StereoCamera> & frame_data)
     dev_residual_img[level].create(mat_size, CV_32FC4);
     dev_residual_img[level].setTo(cv::Scalar(0,0,0,1));
 #else
-    ref_dense_points_[level].create(mat_size.height/EVERY_NTH_PIXEL,
-                                    mat_size.width/EVERY_NTH_PIXEL, CV_32FC4);
-    residual_img[level].create(mat_size.height/EVERY_NTH_PIXEL,
-                               mat_size.width/EVERY_NTH_PIXEL, CV_32FC4);
+    assert(mat_size.width%EVERY_NTH_PIXEL==0);
+    assert(mat_size.height%EVERY_NTH_PIXEL==0);
+    // If this asserts fails, we need to verify whether the logic below still
+    // works, or we need to adapt the code!
+
+    int width_times_factor = mat_size.width/EVERY_NTH_PIXEL;
+    int height_times_factor = mat_size.height/EVERY_NTH_PIXEL;
+
+    ref_dense_points_[level].create(height_times_factor,
+                                    width_times_factor, CV_32FC4);
+    residual_img[level].create(height_times_factor,
+                               width_times_factor, CV_32FC4);
     residual_img[level].setTo(cv::Scalar(0,0,0,1));
 #endif
   }
diff --git a/src/dense_tracking.h b/src/dense_tracking.h
index 00e70a3..76feb82 100644
--- a/src/dense_tracking.h
+++ b/src/dense_tracking.h
@@ -75,6 +75,8 @@ class DenseTracker
 
 private:
   static const int EVERY_NTH_PIXEL = 4;
+  //todo: different factor for different pyramid levels!
+
   const FrameData<StereoCamera> & frame_data_;
 
 #ifdef SCAVISLAM_CUDA_SUPPORT
diff --git a/src/placerecognizer.cpp b/src/placerecognizer.cpp
index 53dcdd8..d75b369 100644
--- a/src/placerecognizer.cpp
+++ b/src/placerecognizer.cpp
@@ -25,6 +25,9 @@
 #include "ransac_models.h"
 #include "ransac.hpp"
 
+// Thanks a lot to Adrien Angeli for all help and discussion concerning
+// place recognition using "bag of words".
+
 namespace ScaViSLAM
 {