Yeah, they say so themselves in the paper: > "In this work, we were more interested in building a proof- of-concept system rather than optimizing for performan

Yeah, they say so themselves in the paper: > "In this work, we were more interested in building a proof- of-concept system rather than optimizing for performance. As a result, our current implementation is slow. It is difficult to measure the exact timings, as we distributed parts of the SfM reconstruction and source selection to multiple machines on a cluster. Table 2 lists an informal summary of our computation times. We expect that substantial speedups are possible by replacing the incremental SfM reconstruction with a real-time SLAM system [Klein and Murray 2009], and finding a faster heuristic for preventing selecting of oc- cluded scene parts in the source selection. We leave these speed-ups to future work."