Human Pose Estimation Method Based on Deformable Convolution
A human body posture and convolution technology, applied in the field of computer vision and pattern recognition, can solve the problems of unrobust estimation performance, immature, and affect accuracy, and achieve the effect of simple deformation and increased accuracy
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0045] In complex scenes, the human body has a special posture or the limbs are distorted. Due to the influence of light refraction or reflection due to environmental reasons, when the scale of the human body in the image changes greatly, the estimation is not accurate enough, and the estimation performance is not robust and immature. Unable to reach application level. The present invention conducts research on these current situations, and proposes a human body pose estimation method based on deformable convolution, see figure 1 , including the following steps:
[0046] (1) Obtain training images:
[0047] (1a) Use the target detection network Mask RCNN to detect the image containing the person, detect the person target, separate the individual person, and return the bounding box of the individual image.
[0048] (1b) Crop the bounding box, obtain the individual image of the person, fill in the constant value around the image to make it into a square image, use it as the tr...
Embodiment 2
[0065] The human body pose estimation method based on deformed convolution is the same as embodiment 1, the deformed convolution module of the deformed convolution kernel described in step 3, and its forward propagation steps are as follows:
[0066] 3.1, Input the input feature map of the deformed convolution module of the deformed convolution kernel into the bias convolution, and obtain the convolution kernel sampling bias feature map output by the bias convolution. The size of the convolution kernel sampling bias feature map should be set is H×W, where H and W are the height and width of the output feature map, respectively, and the number of channels for the bias feature map should be set to 2 k 2 n c , where k is the side length of the convolution kernel, n c is the number of input channels, and the offset feature map contains the offset Δp of the two axes corresponding to the sampling points in each convolution kernel on the feature map in each input channel n .
[00...
Embodiment 3
[0077] The human body pose estimation method based on deformation convolution is the same as embodiment 1-2, the deformation convolution module of the deformation feature map described in step 3, and its forward propagation steps are as follows:
[0078] 3.3, Input the input feature map of the deformed convolution module of the deformed feature map into the bias convolution, and obtain the bias feature map of the input feature map output by the bias convolution. The size of the bias feature map of the input feature map should be set to is H×W, where H and W are the height and width of the input feature map, respectively, and the number of channels for the bias feature map should be set to 2n c , n c is the number of input channels, and the offset feature map contains the offset Δp of the two axes for each point on the feature map in each channel of the input 0 ;
[0079] 3.4, according to the bias Δp in the bias feature map of the input feature map 0 Obtain the deformed con...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More