Efficient Fine-Grained Object Detection
19
Cascade: The method as described in [2] is applied on the baseline architecture,
with the purpose of simultaneously training multiple classifiers optimized at different Interception over Union (IoU) thresholds. Given that each classifier refines
the input regions to align better with their corresponding targets, a sequence of
cooperative classifiers is deployed to progressively increase their quality as well
as the quality of the regions in a single pass.
EA-CNN P : Although the uniform anchoring has proven to be quite effective
in most cases, generating anchors in discrete pixel intervals with fixed shape
and size often results in misalignment between the anchors and their respective
ground truth targets, which can be critical in the case of small objects detection. This can be overcome by generating additional candidate regions through
applying the maximally stable extremal regions (MSER) algorithm on an image
edge-enhanced by simplified Gabor wavelets (SGWs). From this step, only the
small-sized edge information regions, referred to as edge anchors, are considered.
Finally, all the edge anchors are passed to the second stage classifier along with
the regions proposed by the RPN.
EA-CNN C : The integration of the edge anchors into the RPN should retain
the scale-specific feature maps of the FPN and the mapping between bounding
box regressors and identical-shaped anchors. To this end, the edge anchors are
refined to match the closest available shape, size and location configurations,
as dictated by the hyper-parameters of ratio, scale, and input image dimension
respectively. In order to minimize the refinement, additional enlarged feature
maps dedicated to the edge anchors, called edge maps, are introduced, which
correspond to different scales relevant to small objects. After the modifications
described above the RPN is able to evaluate regions given both edge and regular
anchors as input. Based on that, MSER is applied to both grayscale input image
and its edge-enhanced version resulting in more but less precise edge anchors
(Fig. 1).
3 Experiments
Within the context of HR-Recycler project, a new image dataset was created by
capturing multiple PC-Tower devices during their disassembly procedure. The
dataset was appropriately annotated by manually segmenting its components at
various disassembly stages, which was then used for experimental testing.
Implementation Details: The input images are rescaled such as their biggest
side is 512 pixels wide, while their aspect ratio is retained; although feeding
images of a higher resolution would most likely result in better detection performance, the application’s need for real-time operation is limiting. So as to
effectively locate small ground truth targets, we use scales of 10, 12, 15 and
20 pixels with the first one corresponding to the uniformly generated anchors
and the rest of them to the edge maps. When the cascade method is used, the
base classifier is followed by two extra classification heads of 0.55 and 0.60 IoU
thresholds. Finally, for the purpose of generating the edge anchors, the regions
19
Cascade: The method as described in [2] is applied on the baseline architecture,
with the purpose of simultaneously training multiple classifiers optimized at different Interception over Union (IoU) thresholds. Given that each classifier refines
the input regions to align better with their corresponding targets, a sequence of
cooperative classifiers is deployed to progressively increase their quality as well
as the quality of the regions in a single pass.
EA-CNN P : Although the uniform anchoring has proven to be quite effective
in most cases, generating anchors in discrete pixel intervals with fixed shape
and size often results in misalignment between the anchors and their respective
ground truth targets, which can be critical in the case of small objects detection. This can be overcome by generating additional candidate regions through
applying the maximally stable extremal regions (MSER) algorithm on an image
edge-enhanced by simplified Gabor wavelets (SGWs). From this step, only the
small-sized edge information regions, referred to as edge anchors, are considered.
Finally, all the edge anchors are passed to the second stage classifier along with
the regions proposed by the RPN.
EA-CNN C : The integration of the edge anchors into the RPN should retain
the scale-specific feature maps of the FPN and the mapping between bounding
box regressors and identical-shaped anchors. To this end, the edge anchors are
refined to match the closest available shape, size and location configurations,
as dictated by the hyper-parameters of ratio, scale, and input image dimension
respectively. In order to minimize the refinement, additional enlarged feature
maps dedicated to the edge anchors, called edge maps, are introduced, which
correspond to different scales relevant to small objects. After the modifications
described above the RPN is able to evaluate regions given both edge and regular
anchors as input. Based on that, MSER is applied to both grayscale input image
and its edge-enhanced version resulting in more but less precise edge anchors
(Fig. 1).
3 Experiments
Within the context of HR-Recycler project, a new image dataset was created by
capturing multiple PC-Tower devices during their disassembly procedure. The
dataset was appropriately annotated by manually segmenting its components at
various disassembly stages, which was then used for experimental testing.
Implementation Details: The input images are rescaled such as their biggest
side is 512 pixels wide, while their aspect ratio is retained; although feeding
images of a higher resolution would most likely result in better detection performance, the application’s need for real-time operation is limiting. So as to
effectively locate small ground truth targets, we use scales of 10, 12, 15 and
20 pixels with the first one corresponding to the uniformly generated anchors
and the rest of them to the edge maps. When the cascade method is used, the
base classifier is followed by two extra classification heads of 0.55 and 0.60 IoU
thresholds. Finally, for the purpose of generating the edge anchors, the regions