The current version in the repository (latest commit 25668de in Nov. 2015) has merge conflicts in sc_complete.m (<<<<<<< HEAD .... >>>>>>> 80fe71bd858652004119dc8946439915385cc282) and I am not sure why the image filename is also a parameter of the function. A 'cache' and 'result' directories need to be created in the root directory (otherwise it crashes). I also had to download binaries for VLFeat 0.9.20 at http://www.vlfeat.org/download/ since I could not compile it (this library does not seem maintained). The code to detect vanishing points is not provided (the paper just refers to the classical book of Hartley and Zisserman) but instead a binary Windows executable is. Ultimately, the code runs in about 5 min for one image in 3216x2136, or 40s in 640x480. However, it produces results that are very far from the quality of the results shown in the paper. I tried reproducing Fig 8 and obtained the results at https://pasteboard.co/ILTnRz4.jpg (images scaled and cropped to roughly match Fig. 8). I did 3 runs per image to make sure I didn't get unlucky random seeds. A note says this is a matlab reimplementation of the paper (by the paper first author), not the original code used to produce the images though. Also, a todo.txt present in the github seems to indicate there is no spatial coherence cost in this implementation, which could explain the discrepancy between advertised and obtained results, but I am not sure which term exactly this corresponds to in the paper (E_proximity?).
If you want to contribute with another review, please follow these instructions.
Please consider to cut/paste/edit the raw JSON data attached to this paper.