For the CA Tract (challenge dataset), the generated binary envelope image does not fully cover the tract. If some other (bigger) image is selected as reference ( i.e. the complete brain parcellation image), the generated envelopes differ.
It seems, that the geometry (including the bounding box) is not correctly computed from the geometry of the tract.
One effect is caused by implicitly casting the floating-point extent value bbox[2 * i+1] - bbox[ 2*i] to a integer region size ( here at least ceil() should be applied prior to the cast).
The second effect is by the difference in object- and image-geometry ( corner-based vs. center-based), so we need to take the 0.5 offset here into account when creating the image.
both in the itk::TractDensityImageFilter