Window dependent feature regions and strict spatial layout for object detection
View Patent ↗Systems and methods for object detection by receiving an image; segmenting the image and identifying candidate bounding boxes which may contain an object; for each candidate bounding box, dividing the box into overlapped small patches, and extracting dense features from the patches; during a training phase, applying a learning process to learn one or more discriminative classification models to classify negative boxes and positive boxes; and during an operational phase, for a new box generated from the image, applying the learned classification model to classify whether the box contains an object.
1. A method for object detection, comprising:
receiving an image;
segmenting the image and identifying candidate bounding boxes which may contain an object;
for each candidate bounding box, dividing the box into overlapped small patches, and extracting dense features from the patches;
during a training phase, applying a learning process to learn one or more discriminative classification models to classify negative boxes and positive boxes;
during an operational phase, for a new box generated from the image, applying the learned classification model to classify whether the box contains an object, and
generating a weak classifier using a lookup table:
h
(
x
)
=
∑
o
=
1
n
-
1
v
o
1
(
B
(
x
)
=
o
)
,
where h(x) is a piece-wise linear function defined by a lookup table, v o is a table value for an o th entry, B(x) quantizes a feature value x into a table entry, and 1(·) is an indicator function.
2. The method of claim 1 , comprising counting features appearing in candidate sub-regions.
3. The method of claim 1 , comprising:
forming at least an object bounding box for a location; and
applying a cascaded boosting classifier to each object bounding box, with each weak classifier taking a feature response of a region inside the bounding box as its input and then the region is in turn represented by a group of small sub-regions (regionlets).
4. The method of claim 1 , comprising selecting sets of regionlets from a pool of candidate regionlet groups by boosting.
5. The method of claim 1 , comprising using a regionlet representation to model fine-grained spatial appearance layouts.
6. The method of claim 1 , comprising aggregating feature responses of regionlets within one group to a one dimensional feature robust to local deformation.
7. The method of claim 1 , comprising determining a permutation invariant feature operation on features extracted from regionlets as
T
(
R
)
=
∑
j
=
1
N
R
α
j
T
(
r
j
)
,
subject
to
α
j
∈
{
0
,
1
}
,
∑
j
=
1
N
R
α
j
=
1
,
where T (R) as a feature representation for region R, T (r j ) as a feature extracted from the j th regionlet r j in R, N R is a total number of regionlets in region R, α j is a binary variable, either 0 or 1.
8. The method of claim 1 , comprising applying max-pooling over regionlet features as:
T
(
R
)
=
max
j
T
(
r
j
)
.
9. The method of claim 1 , comprising normalizing regionlets using detection windows.
10. The method of claim 1 , comprising normalizing regionlets by by using relative positions and sizes of the regionlets and their groups to an object bounding box.
11. The method of claim 1 , comprising training with boosting regionlet features.
12. The method of claim 1 , comprising using RealBoost to train cascaded classifiers for object detection.
13. A system for object detection, comprising:
a camera;
a processor coupled to the camera;
code for receiving an image;
code for segmenting the image and identifying candidate bounding boxes which may contain an object;
code for each candidate bounding box, dividing the box into overlapped small patches, and extracting dense features from the patches;
code for applying a learning process during a training phase to learn one or more discriminative classification models to classify negative boxes and positive boxes;
code for applying the learned classification model to a new image to classify whether the box contains an object or not, and
generating a weak classifier using a lookup table:
h
(
x
)
=
∑
o
=
1
n
-
1
v
o
1
(
B
(
x
)
=
0
)
,
where h(x) is a piece-wise linear function defined by a lookup table, v o is a table value for an o th entry, B(x) quantizes a feature value x into a table entry, and 1(·) is an indicator function.
14. The system of claim 13 , comprising code for dividing a detection window into small patches.
15. The system of claim 13 , comprising code for defining a patch's size by determining a ratio of the patch's absolute size and a detection window size.
16. The system of claim 13 , comprising code for using raw feature and concatenating the feature together.
17. The system of claim 13 , comprising code for randomly generating a plurality of sub-regions in a candidate region with different sizes and aspect ratios.
18. The system of claim 13 , comprising code for selecting specific feature dimensions in a specific spatial layout for efficient classification.