[spicysharkco]

Fixes #6174.

VoxelGridCovariance::applyFilter already rejects input if the number of voxels would overflow an int, but it only looks at the extent (max_p - min_p) * inverse_leaf_size_. The indices of the first and last voxel, floor (min_p * inverse_leaf_size_) and floor (max_p * inverse_leaf_size_), are then cast to int without a range check. For a cloud far from the origin and a small leaf size they do not fit, which is undefined behavior, and distinct points end up in the same leaf.

Input from the issue: 10 points at x = 1e7 + i, leaf size 1e-3 along x. getLeaves ().size () is 1 instead of 10 (also what I get on arm64). VoxelGrid::applyFilter has the same pattern and returns 1 point for that input, so it is fixed in a second commit.

The fix checks, per axis and in float (the type of the expression that is cast), that both indices fit into an int. If not, it warns with the existing message and returns the same way as the existing check: empty output for VoxelGridCovariance, the unfiltered input for VoxelGrid.

I did not subtract min_p before multiplying, as suggested in the issue. That moves the voxel boundaries from multiples of the leaf size to min_p + k * leaf size, so different points would share a leaf and the NDT results of existing users would change. Rejecting also fits the public interface, which exposes the grid with int coordinates (getMinBoxCoordinates, getGridCoordinates, getCentroidIndexAt). If a correct result for such clouds is preferred, that would be a separate change.

Tests: VoxelGridCovariance.VoxelIndicesBeyondIntRange and VoxelGrid.VoxelIndicesBeyondIntRange in test/filters/test_filters.cpp, along each axis. Offsets 0 and +-1e6 keep working, offsets +-1e7 and clouds where only the smallest or only the largest index is out of range are rejected. Without the fix the rejected cases fail (1 or 2 leaves or points instead of the expected result), with the fix the whole test_filters binary passes.

A third commit clears leaves_ at the start of VoxelGridCovariance::applyFilter. Before, every early return (no input, the existing extent check, and now the index check) left the leaves of the previous input in place, so a reused filter answered getLeaves, getLeaf and getNeighborhoodAtPoint from the old grid. VoxelGridCovariance.RejectedInputClearsLeaves covers that, and VoxelGrid.VoxelIndicesAtIntRangeBounds checks the values at both ends of the int range (2^31 - 128 and -2^31 are accepted, 2^31 and -2^31 - 256 are rejected). I can move the third commit to its own PR if you prefer.

VoxelGrid<PCLPointCloud2> (filters/src/voxel_grid.cpp), UniformSampling and VoxelGridLabel have the same unchecked cast and are not changed here.

[mvieth]

Thanks for the PR! Is it possible to replace these "magic number" with some constant, probably static_cast<float>(std::numeric_limits<int>::lowest()) and static_cast<float>(std::numeric_limits<int>::max()), respectively? I think that would make the intent clearer.

[spicysharkco]

Thanks, @mvieth. Done in both voxel_grid.hpp and voxel_grid_covariance.hpp in 9ece07d. static_cast<float>(std::numeric_limits<int>::max()) is 2^31 as a float, so the < keeps the same exclusive upper bound and the behavior is unchanged.

[mvieth]

Thanks!