A detailed description for pruning used in the inference algorithm.
The proposal A has three children proposal
B1, B2 and B3. The lines between B1, B2, B3 is L1 (B1 and B2), L2 (B1 and B3)
and L3 (B2 and B3)
- Energy function. The threshold is set
using RECALL of each node (Recall = 95%). (For And Node
and Leaf Node)
- The number of proposal of each node. Each
node has at most 200 proposals after NMS. (For And Node
and Leaf Node)
- The scale and rotation of each proposal.
(For And Node and Leaf Node)
- The range of ratio is between 0.5 to
1.5.
- The range of rotation is from -PI/3 to
PI/3.
- Leaf Node: Only using
rotation.
The difference of scale and rotation
between the proposal and its children proposal. (For And
Node)
- Scale: 1.5
- Rotation: PI/4
- If its children
proposals are from leaf nodes, only Rotation is used.
Difference of scale and rotation between
these lines and children proposals are used for pruning. For example, the
difference between B1 and L1, the difference between B2 and L1 and so on. (For
And Node)
- Scale: 1.5
- Rotation: PI/4
- If its children
proposals are from leaf nodes, only Rotation is used.
- The pruning doesn't impact the
performane, however, it will speed up the algorithm by 2 seconds.
The scale and rotation of
L1, L2, L3 (For And Node)
- The range of ratio is between 0.5 to
1.5.
- The range of rotation is from -PI/3 to
PI/3.