Adaptive cross partition for learning weak classifiers
View Patent ↗Systems and methods are disclosed to perform object detection for images from an image sensor by reusing a 1-dimensional feature from a previously learned weak classifier and selecting a new feature to construct a 2-dimensional feature space; and cross partitioning the 2-dimensional space to learn optimal outputs for instances in each domain within a boosting framework.
1. A process to perform object detection for images from an image sensor, comprising:
reusing a 1-dimensional feature from a previously learned weak classifier and selecting a new feature to construct a 2-dimensional feature space; and
cross partitioning the 2-dimensional space to learn optimal outputs for instances in each domain within a boosting framework;
further comprising aggregating the small block-wise regions into larger ones across both dimensions with discriminability; and learning optimal outputs for instances in each block-wise area independently for a boosting operation.
2. The process of claim 1 , further comprising performing independent merging and joint merging that adapt to a distribution change from different perspectives.
3. The process of claim 2 , wherein the independent merging is unsupervised and the joint merging is supervised.
4. The process of claim 1 , further comprising sharing features between neighboring weak classifiers to improve computational efficiency.
5. The process of claim 1 , wherein the small block-wise regions have uniform distribution of training samples weighted by the boosting operation.
6. The process of claim 1 , further comprising performing independent partitioning on a 2-D feature space, each dimension, independently of each other, is evenly divided into a large number of sub-regions of equal width.
7. The process of claim 1 , further comprising performing independent merging where neighboring sub-regions are merged to generate a new partition with approximately uniform distribution of training samples.
8. The process of claim 1 , further comprising joint merging of neighboring columns or rows to shrink the number of sub-regions while maintaining good discriminability of consequent cross partition.
9. The process of claim 1 , further comprising determining a partition θ that maximizes a Bhattacharyya distance
Bha
(
θ
)
=
∑
j
W
j
+
W
j
-
where
W
j
b
=
∑
k
:
y
k
=
b
,
θ
(
x
k
)
=
j
w
k
is the sum of weights of training samples of category b falling to a j-th sub-region (b=±1).
10. A system to perform object detection for images from
an image sensor, comprising:
means for reusing a 1-dimensional feature from a previously learned weak classifier and selecting a new feature to construct a 2-dimensional feature space;
means for cross partitioning the 2-dimensional space to obtain a number of small block-wise regions with uneven width or height and approximately uniform distribution of training samples;
means for aggregating the small block-wise regions into larger ones across both dimensions with discriminability; and
means for learning optimal outputs for instances in each block-wise area independently for a boosting operation.
11. The system of claim 10 , further comprising means for performing independent merging and joint merging that adapt to a distribution change from different perspectives.
12. The system of claim 11 , wherein the independent merging is unsupervised and the joint merging is supervised.
13. The system of claim 10 , further comprising means for sharing features between neighboring weak classifiers to improve computational efficiency.
14. The system of claim 10 , wherein the small block-wise regions have uniform distribution of training samples weighted by the boosting operation.
15. The system of claim 10 , further comprising means for performing independent partitioning on a 2-D feature space, each dimension, independently of each other, is evenly divided into a large number of sub-regions of equal width.
16. The system of claim 10 , further comprising means for performing independent merging where neighboring sub-regions are merged to generate a new partition with approximately uniform distribution of training samples.
17. The system of claim 10 , further comprising means for combining independent partitions of a plurality of dimensions to achieve a cross partition of the 2-dimensional space.
18. The system of claim 10 , further comprising means for joint merging of neighboring columns or rows to shrink the number of sub-regions while maintaining good discriminability of consequent cross partition.
19. The system of claim 10 , further comprising means for determining a partition θ that maximizes a Bhattacharyya distance
Bha
(
θ
)
=
∑
j
W
j
+
W
j
-
where
W
j
b
=
∑
k
:
y
k
=
b
,
θ
(
x
k
)
=
j
w
k
is the sum of weights of training samples of category b falling to a j-th sub-region (b=±1).