search

AndreiBarsan / DynSLAM

Master's Thesis on Simultaneous Localization and Mapping in dynamic environments. Separately reconstructs both the static environment and the dynamic objects from it, such as cars.
BSD 3-Clause "New" or "Revised" License
576 stars 179 forks source link

Fix depth filtering when using depth maps from dispnet #10

Closed AndreiBarsan closed 7 years ago

AndreiBarsan commented 7 years ago

This issue does not occur when using ELAS, just dispnet; the filterDepth_device kernel fails (but without crashing the rest of the program) at the line tmpz = imageData_in[(x + j) + (y + i) * imgDims.x];

This may also be the root cause of the "floating" artifacts produced when using dispnet depth but not elas depth. Solving this would allow me to turn down the min depth back to 0.5 or so, yay!

Once again, cuda-memcheck is the real MVP.

AndreiBarsan commented 7 years ago

Yep, turns out the ITM code had an off-by-one error which was masked by the fact that they were using VGA input, whose dimensions were a multiple of 16, the block size used for the bilateral filtering.

Once you plugged in arbitrary values for the depth size, reads would start to happen from outside the array bounds, leading to undefined behavior, and, therefore occasional artifacts. This has been fixed in commit AndreiBarsan/InfiniTAM@9211e4f491ea321159a32ad76aa67cd482d18b80.

Cuda-memcheck was VERY helpful!