In 2D physics, it depends on the perspective:
* When viewed from the side, things collide on a plane parallel to the background.
* When viewed from above, things collide on the ground plane.

To determine the correct bounding rectangle, you can imagine a suitable bounding box in 3D, and then project that into the collision plane.

A common analogy for protecting a 3D object onto a 2D plane is shadow casting. If you imagine the plane as the ground, then the projection is the object's shadow when the sun is directly overhead.

To be frank, whether it's intentional or not, the perspective of your game's sprites screams sidescroller; especially in the second image where you appear to be standing on a surface. In this case, the bounds for both objects are the result of casting their shadows onto the background; it would match their apparent size.

In the case of a top-down or isometric perspective, the shadows are cast onto the ground. In true top-down (bird's eye view, where players are a head with shoulders), this matches the apparent size. In isometric, the bounding rectangle would be centered at the base and extend according to how wide the object is when viewed from above.

Of course, this is assuming you're going for "realistic" bounding boxes at all. It's entirely up to you as the game designer whether to use "tight" bounds that accurately reflect the size of the sprite. Many games prefer to extend bounding rectangles (or at least a separate "hitbox") for enemies so that they're easier to hit.

As for layering, try this out:

*Attack on Titan theme begins*

lol god damnit.