Object detection with Regionlets re-localization
View Patent ↗An object detector includes a bottom-up object hypotheses generation unit; a top-down object search with supervised descent unit; and an object re-localization unit with a localization model.
1. An object detector, comprising:
a computer to perform the following units:
a bottom-up object hypotheses generation unit;
a top-down object search with supervised descent unit;
an object re-localization unit with a localization model; and
an object relocalization unit with a localization module that determines:
min
V
{
V
2
+
C
∑
m
=
1
M
max
(
0
,
Δ
L
m
-
V
T
R
m
-
ɛ
)
2
}
,
where V is a coefficient vector to be learned, ΔL m is a normalized localization error of training sample m, R m is a feature extracted from all the Regionlets in an object detection model for the m th sample as explained in the following, M is the total number of training examples, C is a trade-off factor between a regularization and the sum of squared error, ε is a tolerance factor.
2. The detector of claim 1 , comprising a feature extractor that extracts features from the image, and a learning module to train the object detector, wherein the learned object detector is applied to each possible locations to detect the object.
3. The detector of claim 1 , wherein the object detector makes a b binary decision on whether the location presents an object based on scores provided by the object detector.
4. The detector of claim 1 , comprising a location regression module to improve localization of the object.
5. The detector of claim 1 , comprising Regionlets having extracted features into a multi-dimensional binary vector and wherein binary vectors obtained from all Regionlets features are concatenated for regression training.
6. The detector of claim 1 , comprising a least square learning module to learn the coefficients for location regression based on the binary vector.
7. The detector of claim 1 , wherein the object hypotheses are formed through objectness measurement, saliency analysis or their combinations, or segmentation cues.
8. The detector of claim 1 , wherein the bottom-up object hypotheses generation unit splits the object location space evenly to avoid the search algorithm converging to the same local minimum.
9. The detector of claim 1 , comprising a low-level segmentation unit to propose the object hypotheses.
10. The detector of claim 1 , comprising a superpixel segmentation unit to merge similar pixels locally into disjoint sets.
11. The detector of claim 1 , wherein over segmented superpixels are gradually merged to produce larger candidates.
12. The detector of claim 1 , wherein the detection with location relaxation takes coarse detection results and searches the object location guided by discriminatively learned descent mode.
13. The detector of claim 1 , comprising learned supervised descent model is used to predict the next more accurate object location to explore based on observations from the current location.
14. The detector of claim 1 , comprising a segmentation based bottom-up module to generate an initial set of candidate searching locations.
15. The detector of claim 1 , comprising a module to receive over-segments or superpixels of an image and then hierarchically group these small regions to generate object hypotheses and to generate superpixel segments.
16. The detector of claim 15 , wherein a segmented region r i is described by a plurality of characteristics including size of the region (total number of pixels), color histograms, and texture information or gradient orientation histograms.
17. The detector of claim 16 , wherein four neighbor region similarities are defined based on these characteristics as shown in the following equations:
S
c
(
r
i
,
r
j
)
=
∑
k
=
1
n
min
(
c
i
k
,
c
j
k
)
,
where c i k is the k th dimension of the color histogram, sz(r i ) is the number of pixels in image region r i , im stands for the whole image, t i k is the k th dimension of the texture histogram, bb ij is the rectangular region which tightly bound region r i and r j . S c , S s and S t are the color similarity, size similarity, texture similarities, respectively. S f measures how the combined two regions will occupy the rectangular bounding box which tightly bounds them.