The following tutorial shows how to make a reconstruction of one of the training scenes with COLMAP. Furthermore, we show how to align the reconstruction with the ground truth and how to get the precision, recall and F-score. At the end, we show a few examples on how to convert the different SfM files into the *.log file format which will be needed when you do a submission on the test set.

Let's start with the Ignatius dataset. Download the image set or the video from the download page. For the image set, unzip the package and go to step 2. For the video, use the following script to sample key frames from the video (These tools are required: * ffmpeg*,

Get jpg images from the video:> ffmpeg -i Ignatius.mp4 -q:v 1 -vf fps=1 Ignatius/%06d.jpgResize:> mogrify -quality 100 -resize 50% -path Ignatius2/ Ignatius/*.jpgWrite EXIF data:> exiftool -overwrite_original -FocalLength="21" -focallengthin35mmformat="21" -make="Sony" -model="A7SM2" Ignatius2/*.jpg

Download and install COLMAP. Follow the installation guidelines and the COLMAP tutorial to make a reconstruction of our Ignatius2 image set from step 1. Alternatively, you can run this script which includes all necessary steps of the COLMAP pipeline to produce a reconstruction:

> sh get_colmap_reconstruction.sh [COLMAP-exe-directory] [image-set-directory] [project-directory]

After a successful run, the reconstruction "fused.ply" can be found in the [project-directory]/dense folder. If you are working on a training dataset, follow step 3 and step 4 to perform an evaluation with the ground-truth geometry. If you are working on a testing dataset, jump to step 5 to prepare a *.log file and go to the submission page for a submission. However, do reconstruct all models in either the intermediate or the advanced set before submission.

If you are working with a training dataset, download the Ignatius ground-truth file. Since the COLMAP reconstruction has arbitrary scale and orientation, we need to exactly align the reconstruction with the ground truth. This can be done by using MeshLab or CloudCompare. Make sure you set the ground-truth file as a reference, and align the reconstruction to it. You can find tutorials on how to do that on-line. We will have a dedicated library to help performing these task ready soon.

We will soon have a dedicated library to easily perform the necessary tasks for this step. Otherwise you can follow these steps to get an F-score evaluation result yourself:

Let \( \mathcal{G} \) be the ground truth and \( \mathcal{R} \) a reconstructed point set being evaluated. For a reconstructed point \( \mathbf{r} \in \mathcal{R} \), its distance to the ground truth is defined as: \begin{equation} e_{\mathbf{r} \to \mathcal{G}} = \min_{\mathbf{g} \in \mathcal{G}} |\mathbf{r} - \mathbf{g}| . \end{equation} These distances can be aggregated to define the precision of the reconstruction \( \mathcal{R}\) for any distance threshold \( d \): \begin{equation} P(d) = \frac{100}{|\mathcal{R}|} \sum_{\mathbf{r} \in \mathcal{R}} \big[ e_{\mathbf{r} \to \mathcal{G}} < d \big] \label{eq:precision} \end{equation} where \( [\cdot]\) is the Iverson bracket. \( P(d) \) is defined to lie in the range [0,100] and can be interpreted as a percentage. Similarly, for a ground-truth point \( \mathbf{g} \in \mathcal{G} \), its distance to the reconstruction is defined as \begin{equation} e_{\mathbf{g} \to \mathcal{R}} = \min_{\mathbf{r} \in \mathcal{R}} |\mathbf{g} - \mathbf{r}| . \end{equation} The recall of the reconstruction \( \mathcal{R}\) for a distance threshold \( d \) is defined as \begin{equation} R(d) = \frac{100}{|\mathcal{G}|} \sum_{\mathbf{g} \in \mathcal{G}} \big[ e_{\mathbf{g} \to \mathcal{R}} < d \big]. \label{eq:recall} \end{equation} Precision and recall can be combined in a summary measure, the F-score: \begin{equation} F(d) = \frac{2P(d)R(d)}{P(d)+R(d)}. \label{eq:f-score} \end{equation} The F-score at a given threshold \( d \) is the harmonic mean of precision and recall at this threshold. It has the property that if either \( {P(d)\rightarrow 0}\) or \( {R(d)\rightarrow 0}\), then \( {F(d)\rightarrow 0}\). It is thus a better summary measure than the arithmetic mean, which does not have this property.

For a test-set evaluation, the camera poses need to be stored in *.log format. An example of a log file can be seen here. Every input image needs to have an entry in the log file, regardless if it is recognized as a valid view in the SfM bundle or not. **It is okay to have some inaccurate camera poses.** The back-end of our evaluation server is robust to such cases. However, we do recommend to estimate the camera pose as good as possible.
To convert the COLMAP SfM data to the log format you can use this python script and follow the instructions below. The script can also be used to convert MVE or VisualSfM SfM data.

> python convert_to_logfile.py [COLMAP SfM file] [output log-filename] [folder of the input images] [COLMAP/MVE/VSFM] [image format]Example to get the *.log file of the COLMAP reconstruction of this tutorial:> python convert_to_logfile.py Ignatius_COLMAP/sparse/0/ cameras.log Ignatius_COLMAP COLMAP jpgExample to get the *.log file of a MVE based SfM file:> python convert_to_logfile.py mve-folder/synth_0.out mve.log mve_test/views/ MVEExample to get the *.log file of a VisualSfM based SfM file:> python convert_to_logfile.py vsfm-folder/result.nvm vsfm_log.log vsfm_test/ input_images/ VSFM jpg

If you convert COLMAP SfM data you will need the COLMAP read_model.py script in the same folder you are executing the this script from.

If you used the video and sampled your own set of images, it is important to submit a log file which includes interpolated camera poses for all the individual frames of the video. First, create a log file (cameras.log) using the script in Case 1. Then edit a mapping.txt file that includes information about which video frames you sampled. For this tutorial example it would look like the following:

[number of input camera poses] [number of frames in the video] [Image ID] [video frame ID] : (repeat for every image)Example mapping file of this tutorial:263 7845 0 1 1 31 2 61 3 91 4 121 ...

The full mapping file for this tutorial can be downloaded here. Using the log file interpolation function from our toolbox you can now generate a log file (Ignatius.log) which includes interpolated camera poses for every frame:

> python TanksAndTemples/python_toolbox/interpolate_log_file.py cameras.log Ignatius_mapping.txt Ignatius.logIf Ignatius would be part of our test set, you could just take the Ignatius.log file and rename the reconstruction to "Ignatius.ply" and do this for all other reconstructions of the advanced or intermediate set, and you would be ready to make a submission. Check the Submit section of this site for further details on how to proceed with a test set submission.