BVH collision traversal algorithm not looking within each descendant

Question

I'm looking at this code from Game Physics Engine Development for a BVH traversal algorithm, specifically getPotentialContacts and getPotentialContactsWith at the end of the file.

By the looks of this algorithm, it'll compare an initial pair of siblings, but it won't look for collisions within each descendant.

I can't see how this would work on a graph like this one, where dotted lines represent branches, solids are leaf nodes, and the tree depths are represented by spectrum colors (red, orange, yellow, green):

What is it that I'm not understanding here? Do I need another algorithm to find all the contacts within a tree?

I also tried traversing down each of the leafs, but then I end up detecting the collisions twice in many cases -- so that's not it either.

Answer 1

I was having the same problem with this function... so I tried a few approaches and I ended up with this:

  private  int getPotentialContactsWith(
         BVHNode other,
        Vector<PotentialContact> contacts,boolean descend) {

      int count=0;
      //System.out.println(id+" comparando com "+other.id+" contacts.size:"+contacts.size());
      checks++;
      // Early out if we don't overlap or if we have no room
      // to report contacts

      if ((descend) && (!isLeaf())) {
         count += children[0].getPotentialContactsWith(
               children[1], contacts,descend);
      }

      if ((descend) && (!other.isLeaf())) {
         count += other.children[0].getPotentialContactsWith(
                other.children[1], contacts,descend);
       }


        if(!overlaps(other)) return 0;




      // If we're both at leaf nodes, then we have a potential contact
      if (isLeaf() && other.isLeaf())
      {
          if (!alreadyInside(body,other.body,contacts)){
              PotentialContact contact=new PotentialContact(body,other.body);
              contacts.add(contact);} else {errors++;}
          return 1;
      }


      // Determine which node to descend into. If either is
      // a leaf, then we descend the other. If both are branches,
      // then we use the one with the largest size.
      if (other.isLeaf() ||
          (!isLeaf() && volume.getSize() >= other.volume.getSize()))
      {
          // Recurse into ourself
          count += children[0].getPotentialContactsWith(
              other, contacts,false
              );

          // Check we have enough slots to do the other side too
              count += children[1].getPotentialContactsWith(
                  other, contacts,false );
      }
      else
      {
          // Recurse into the other node
          count += getPotentialContactsWith(
              other.children[0], contacts,false);

          // Check we have enough slots to do the other side too
              count += getPotentialContactsWith(
                  other.children[1], contacts,false
                  );

      }


      return count;
    }

//// About the code: Well, its in java but I guess its easy to translate or understand what it is doing. I created the function "alreadyInside" to check if I already added the potential contact and I was increasing the variable errors if I did, so far this code is not adding any repeated potential contact (errors=0), so I will problably drop this function to optimize the code. Also, I added the "descend" parameter, which is a flag which tells when to go further down in the structure.