Hi, thanks a lot for the work! I have a more speculative question, thanks in advance :)
Since Mast3r shows such good performance on ultra wide baseline image pairs, I was also wondering if the network could be re-purposed for loop-closure detection / place recognition (in robotics). A simple idea would be to use the descriptors computed for each image (H x W x 24) and aggregate them in some manner to form a global descriptor, after which NN matching can be performed as per usual.
However, the current network is set-up with two images as input (which of course makes sense). Would you forsee sucess in inputting the same image twice to extract a 'global descriptor'?
Hi, thanks a lot for the work! I have a more speculative question, thanks in advance :)
Since Mast3r shows such good performance on ultra wide baseline image pairs, I was also wondering if the network could be re-purposed for loop-closure detection / place recognition (in robotics). A simple idea would be to use the descriptors computed for each image (H x W x 24) and aggregate them in some manner to form a global descriptor, after which NN matching can be performed as per usual.
However, the current network is set-up with two images as input (which of course makes sense). Would you forsee sucess in inputting the same image twice to extract a 'global descriptor'?