Sparse region-of-interest pooling for object detection
View Patent ↗An image data is convolved with one or more kernels and corresponding one or more feature maps generated. Region of interest maps are extracted from the one or more feature maps, and pooled based on one or more features selected as selective features. Pooling generates a feature vector with dimensionality less than a dimensionality associated with the one or more feature maps. The feature vector is flattened and input as a layer in a neural network. The neural network outputs a classification associated with an object in the image data.
1. A computer-implemented method of detecting an object, comprising:
receiving image data;
convolving the image data with a kernel, the convolving generating a corresponding feature map;
extracting a region of interest map from the feature map;
pooling one or more features selected as one or more selective features in the region of interest map without pooling the region of interest map in entirety, the pooling generating a feature vector with dimensionality less than a dimensionality associated with the region of interest map;
flattening the feature vector and inputting the flattened feature vector as a layer in a neural network; and
outputting by the neural network a classification associated with the object,
wherein the one or more selective features are selected by executing a statistical algorithm on data representing classes of objects and ranking features in the region of interest map, wherein the one or more selective features are selected from the ranked features in the region of interest map based on meeting a threshold.
2. The method of claim 1 , wherein the neural network comprises a fully connected multi-layer perceptron neural network.
3. A computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions readable by a device to cause the device to:
receive image data;
convolve the image data with a kernel, the convolving generating a corresponding feature map;
extract a region of interest map from the feature map;
pool one or more features selected as one or more selective features in the region of interest map without pooling the region of interest map in entirety, the pooling generating a feature vector with dimensionality less than a dimensionality associated with the region of interest map;
flatten the feature vector and inputting the flattened feature vector as a layer in a neural network; and
output by the neural network a classification associated with the object,
wherein the one or more selective features are selected by executing a statistical algorithm on data representing classes of objects and ranking features in the region of interest map, wherein the one or more selective features are selected from the ranked features in the region of interest map based on meeting a threshold.
4. The computer program product of claim 3 , wherein the pooling further comprises sub-sampling the feature vector by sampling image pixels from a subset of locations of the feature vector.
5. The method of claim 1 , wherein the neural network is trained to output one versus rest classification.
6. The method of claim 1 , wherein the statistical algorithm comprises at least one of least absolute shrinkage and selection operator (LASSO), sparse encoding, and regularized sparsity.
7. The method of claim 1 , wherein the one or more selective features are selected offline.
8. The method of claim 1 , wherein the region of interest map has dimensionality less than the feature map.
9. The method of claim 1 , wherein the pooling further comprises sub-sampling the feature vector by sampling image pixels from a subset of locations of the feature vector.
10. The computer program product of claim 3 , wherein the neural network comprises a fully connected multi-layer perceptron neural network.
11. The computer program product of claim 3 , wherein the statistical algorithm comprises at least one of least absolute shrinkage and selection operator (LASSO), sparse encoding, and regularized sparsity.
12. The computer program product of claim 3 , wherein the one or more selective features are selected offline.
13. The computer program product of claim 3 , wherein the region of interest map has dimensionality less than the feature map.
14. A system of detecting an object, comprising:
a hardware processor coupled with a memory device,
the hardware processor operable to at least:
receive image data;
convolve the image data with a kernel, the convolving generating a corresponding feature map;
extract a region of interest map from the feature map;
pool one or more features selected as one or more selective features in the region of interest map without pooling the region of interest map in entirety to generate a feature vector with dimensionality less than a dimensionality associated with the region of interest map;
flatten the feature vector and inputting the flattened feature vector as a layer in a neural network; and
output by the neural network a classification associated with the object,
wherein the one or more selective features are selected by executing a statistical algorithm on data representing classes of objects and ranking features in the region of interest map, wherein the one or more selective features are selected from the ranked features in the region of interest map based on meeting a threshold.
15. The system of claim 14 , wherein the neural network comprises a fully connected multi-layer perceptron neural network.
16. The system of claim 14 , wherein the neural network is trained to output one versus rest classification.
17. The system of claim 14 , wherein the statistical algorithm comprises at least one of least absolute shrinkage and selection operator (LASSO), sparse encoding, and regularized sparsity.