Method, systems, and apparatuses for heatmap regression in a neural network
Methods, systems, and apparatuses for heatmap regression for neural networks are provided, including for systems with restricted computational resources, such as embedded systems. An exemplary method may include utilizing a trained neural network having a plurality of pooling layers, a plurality of convolutional layers, and an output layer to generate a heatmap. The plurality of pooling layers may reduce the spatial resolution of an input image, though the resolution may not be reduced below the spatial resolution of the output layer. The plurality of convolutional layers may utilize a plurality of depthwise convolutions and pointwise convolutions that will be stacked to provide a plurality of bottlenecks.
1 . A method for heatmap regression with a neural network comprising:
receiving an image;
generating a heatmap from the image with the neural network by:
reducing a spatial resolution of the image with one or more pooling layers, wherein a reduction in the spatial resolution is to a spatial resolution of an output layer;
performing a plurality of convolution operations with a plurality of convolution layers, wherein the plurality of convolution layers each have the same spatial resolution, and wherein the neural network includes a plurality of depthwise convolution and pointwise convolution bottlenecks varying a plurality depth channels associated with the plurality of convolution layers;
generating the output layer with a depth channel of a first depth resolution; and
generating a heatmap with a plurality of key points, wherein a number of the plurality of key points equals the first depth resolution.
2 . The method for heatmap regression with a neural network of claim 1 , wherein a depth channel resolution of at least one of the plurality of depthwise convolution and pointwise convolution bottlenecks has the same depth channel resolution as the output layer.
3 . The method for heatmap regression with a neural network of claim 1 , wherein a depth channel resolution of each of the plurality of depthwise convolution and pointwise convolution bottlenecks is the same.
4 . The method for heatmap regression with a neural network of claim 1 , wherein a first depth channel resolution of a first bottleneck of the plurality of depthwise convolution and pointwise convolution bottlenecks is different than a second depth channel resolution of a second bottleneck of the plurality of depthwise convolution and pointwise convolution bottlenecks.
5 . The method for heatmap regression with a neural network of claim 1 , wherein the one or more pooling layers include two pooling layers.
6 . The method for heatmap regression with a neural network of claim 1 , wherein the spatial resolution of the output layer is 64×64 pixels and a depth channel resolution of the output layer is 17 features.
7 . The method for heatmap regression with a neural network of claim 1 , wherein the plurality of key points of the heatmap are coordinates in the image identifying a person.
8 . A system for heatmap regression with a neural network comprising:
at least one processor and at least one memory coupled to the at least one processor, wherein the at least one processor is configured to:
receive an image;
generate a heatmap from the image with the neural network by:
reduce a spatial resolution of the image with one or more pooling layers, wherein a reduction in the spatial resolution is to a spatial resolution of an output layer;
perform a plurality of convolution operations with a plurality of convolution layers, wherein the plurality of convolution layers each have the same spatial resolution, and wherein the neural network includes a plurality of depthwise convolution and pointwise convolution bottlenecks varying a plurality depth channels associated with the plurality of convolution layers;
generate the output layer with a depth channel of a first depth resolution; and
generate a heatmap with a plurality of key points, wherein a number of the plurality of key points equals the first depth resolution.
9 . The system for heatmap regression with a neural network of claim 8 , wherein a depth channel resolution of at least one of the plurality of depthwise convolution and pointwise convolution bottlenecks has the same depth channel resolution as the output layer.
10 . The system for heatmap regression with a neural network of claim 8 , wherein a depth channel resolution of each of the plurality of depthwise convolution and pointwise convolution bottlenecks is the same.
11 . The system for heatmap regression with a neural network of claim 8 , wherein a first depth channel resolution of a first bottleneck of the plurality of depthwise convolution and pointwise convolution bottlenecks is different than a second depth channel resolution of a second bottleneck of the plurality of depthwise convolution and pointwise convolution bottlenecks.
12 . The system for heatmap regression with a neural network of claim 8 , wherein the one or more pooling layers include two pooling layers.
13 . The system for heatmap regression with a neural network of claim 8 , wherein the spatial resolution of the output layer is 64×64 pixels and a depth channel resolution of the output layer is 17 features.
14 . The system for heatmap regression with a neural network of claim 8 , wherein the pluality of key points of the heatmap are coordinates in the image identifying a person.
15 . A computer program product for heatmap regression with a neural network comprising at least one non-transitory computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions comprising an executable portion configured to:
receive an image;
generate a heatmap from the image with the neural network by:
reduce a spatial resolution of the image with one or more pooling layers, wherein a reduction in the spatial resolution is to a spatial resolution of an output layer;
perform a plurality of convolution operations with a plurality of convolution layers, wherein the plurality of convolution layers each have the same spatial resolution, and wherein the neural network includes a plurality of depthwise convolution and pointwise convolution bottlenecks varying a plurality depth channels associated with the plurality of convolution layers;
generate the output layer with a depth channel of a first depth resolution; and
generate a heatmap with a plurality of key points, wherein a number of the plurality of key points equals the first depth resolution.
16 . The computer program product for heatmap regression with a neural network of claim 15 , wherein a depth channel resolution of at least one of the plurality of depthwise convolution and pointwise convolution bottlenecks has the same depth channel resolution as the output layer.
17 . The computer program product for heatmap regression with a neural network of claim 15 , wherein a depth channel resolution of each of the plurality of depthwise convolution and pointwise convolution bottlenecks is the same.
18 . The computer program product for heatmap regression with a neural network of claim 15 , wherein a first depth channel resolution of a first bottleneck of the plurality of depthwise convolution and pointwise convolution bottlenecks is different than a second depth channel resolution of a second bottleneck of the plurality of depthwise convolution and pointwise convolution bottlenecks.
19 . The computer program product for heatmap regression with a neural network of claim 15 , wherein the one or more pooling layers include two pooling layers.
20 . The computer program product for heatmap regression with a neural network of claim 15 , wherein the plurality of key points of the heatmap are coordinates in the image identifying a person.