Calculation by chunk is useful on very large projects, but unfortunately if you want to merge tie points on large projects, memory management is catastrophic (out of memory with 256GB available). The steps look identical to the alignment, so why is there no fine task subdivision like with all the other processes? This would make it possible to manage large projects. Computing time is rarely a problem because the machines can be left to run on their own. What is problematic is stability. I no longer count the number of crashes encountered after 5-6 hours or more of calculation. It's wasted time.
The method also seems weird, the images are duplicated, and this causes problems later when calculating the depth (the identical images are no longer exactly in the same place, which creates an infinite depth). What we want after having aligned the chunks by images, it is simply that the identical images are merged and their correspondences also (either in an additive way for the couples being in the two chunks, or by recreating the correspondences only for the couples present in the two chunks, not for everything!).
Topic: Merge chunks: resources consumption and method (Read 1557 times)
