The core implementation of "Cascaded Parallel Filtering for Memory-Efficient Image-Based Localization" (ICCV 2019), by Wentao Cheng, Weisi Lin, Kan Chen and Xinfeng Zhang. ------------ Requirements ------------ Our code is based on the ACG_Localizer project by Prof Torsten Sattler. Note that we use this project for convenience only. The following libraries are required: * FLANN. The Fast Library for Approximate Nearest Neighbors written Marius Muja and David G. Lowe, available at http://people.cs.ubc.ca/~mariusm/index.php/FLANN/FLANN . ACG Localizer has been tested with version 1.6.11. Due to an API breaking change in version 1.7.0, we suggest that you use 1.6.11, which (as of 02/29/2012) can still be found at http://people.cs.ubc.ca/~mariusm/uploads/FLANN/flann-1.6.11-src.zip We plan to update our source code to be compliant with newer versions of FLANN. * jhead A tool to read exif tags from jpg files, available at http://www.sentex.net/~mwandel/jhead/index.html . It is already included in this release of ACG Localizer. * CMake (http://www.cmake.org/) * Eigen 3.2.0 or newer (http://eigen.tuxfamily.org/) ------------ Installation ------------ Assuming that you unpacked Cascade_Parallel_Filtering into the directory SOME_DIRECTORY/Cascade_Parallel_Filtering, you first have to edit the CMake files that are responsible for finding the required libraries. These files are located in SOME_DIRECTORY/Cascade_Parallel_Filtering/cmake. You have to edit the FindFLANN.cmake files to fill in the directories where the include and library files of both libraries are located. cmake: * cd SOME_DIRECTORY/Cascade_Parallel_Filtering * mkdir build * cd build * cmake .. If no library is missing and if CMake threw no errors you can build the software by simply typing * make To get the best performance, make sure that you set the build type to "Release". Typing 'make install' will install the generated executables in SOME_DIRECTORY/Cascade_Parallel_Filtering/build/bin . Notice that to achieve timing results similar to those in the paper, you have to set the build type of CMake to "Release". ----- Usage ----- In this code, we use the .info format to represent the 3D SfM models. For the meanings of parameters, please refer to the descriptions in the code. We show a localization pipeline on the Aachen Day-Night dataset as an example. Step 1: You need to use the ./compute_desc_assignments to quantize feature descriptors in a SfM model into a visual vocabulary (either in integer mean per visual word or all descriptors) ./compute_desc_assignments aachen_cvpr2018_db.info 1 10000 aachen_cvpr2018_10k.txt aachen_10k_cvpr2018_branch100_mean.bin 6 1 1 100 1 Step 2: After quantization, you need to transfer the SIFT/RootSIFT descriptors into binary signatures. Following is the example: ./compute_hamming_threshold 10000 aachen_cvpr2018_10k.txt aachen_10k_cvpr2018_branch100_mean.bin hamming_projection_matrix.txt aachen_10k_cvpr2018_branch100_mean_hamming_threshold.txt Step 3: Then you can use our cascaded_parallel_filtering as following ./cascaded_parallel_filtering_aachenDayNight day_time_queries_with_intrinsics.txt 10000 aachen_cvpr2018_10k.txt aachen_10k_cvpr2018_branch100_mean_hamming_threshold.txt 100 aachen_cvpr2018_db.info 16 3 20 50 0.3 0.8 output/aachen_cvpr_10k_2d_day.txt output/aachen_cvpr_10k_3d_day.txt ./cascaded_parallel_filtering_aachenDayNight night_time_queries_with_intrinsics.txt 10000 aachen_cvpr2018_10k.txt aachen_10k_cvpr2018_branch100_mean_hamming_threshold.txt 100 aachen_cvpr2018_db.info 16 3 20 50 0.3 0.8 output/aachen_cvpr_10k_2d_night.txt output/aachen_cvpr_10k_3d_night.txt Note: The following step is not included in this repository due to some compatibility issues. In addition, we recommend you to use some RANSAC variants, e.g., LO-RANSAC, instead of the standard RANSAC scheme used in our paper. For the request of code, please contact wcheng005@e.ntu.edu.sg Step 4: The above program will generate two output files. One file stores the 2D positions of matches (first the matches for computing the auxiliary camera pose, second serve as visibility-wise match pool). In general, you have the following two options: a: suppose the focal length is unknown, you should use a p4p solver to compute the auxiliary camera pose, and filter the visibility-wise match pool with 10 pixel re-projection error threshold. After you obtain the final matches, you only need to run a p3p solver with voted focal length value for efficiently obtaining the final camera pose. b: suppose camera calibration is fully known, you can simply apply a p3p solver for both geometry-wise filtering and obtaining the final camera pose.