#Running#

• Linux machine with CUDA
• Torch (http://torch.ch/docs/getting-started.html) with OpenCv bindings (https://github.com/VisionLabs/torch-opencv)
• Put the right path to the Locate dataset in testLib.lua at L5.
• Example command: CUDA_VISIBLE_DEVICES=0 th testKeypoints.lua -s kanzelwandbahn

#Notes#

• I hope you have some experience with Lua/Torch. Otherwise we can talk. If you run on Python, one can bridge the gap with https://github.com/hughperkins/pytorch .
• My code detects keypoints in images (opencv), extracts patches around them, computes matching score between them and evaluates correctness with average precision (correct = best match within 15px of the ground truth).
• Input images are loaded in loadImagePair() (one real (3 channels) and one synthetic (17 channels, various shadings)) and then network in loadNetwork().
• extractKeypoints() crops patches around keypoints.
• fullMatch() computes pairwise matches between all synthetic and real keypoints (quadratic number of evaluations). The network returns a score for each pair of patches.