BVHes handle sparsity and very large scenes with variable density very well, so they are good to be able to see which objects you might intersect. kd-Trees are

BVHes handle sparsity and very large scenes with variable density very well, so they are good to be able to see which objects you might intersect. kd-Trees are much better for objects because they handle the high density of geometry very well and are thus very effective if you want to test against a million triangles in close proximity, for example.