System and method for occluding contour detection
A system and method for occluding contour detection using a fully convolutional neural network is disclosed. A particular embodiment includes: receiving an input image; producing a feature map from the input image by semantic segmentation; learning an array of upscaling filters to upscale the feature map into a final dense feature map of a desired size; applying the array of upscaling filters to the feature map to produce contour information of objects and object instances detected in the input image; and applying the contour information onto the input image.
1. A system comprising:
a data processor; and
an occluding object contour detection processing module, executable by the data processor, the occluding object contour detection processing module being configured to at least:
receive an input image;
produce a feature map from the input image;
apply a range of different dilation rates to the feature map to produce a final feature map maintaining a resolution corresponding to training labels, wherein the different dilation rates are applied to each of a plurality of convolution layers, the different dilation rate not having a common factor relationship other than the number one;
match object shapes from the training labels to objects and object instances detected in the input image;
generate, based on the object shapes, contour information of the objects and object instances detected in the input image; and
apply the contour information onto the final feature map.
2. The system of claim 1 wherein the feature map is produced from the input image by semantic segmentation, wherein the semantic segmentation is machine learnable.
3. The system of claim 1 wherein the feature map is produced from the input image by semantic segmentation, wherein the semantic segmentation is performed by a deep convolutional neural network trained on a dataset configured for a traffic environment.
4. The system of claim 1 being configured to operate within a fully convolutional network.
5. The system of claim 1 wherein the contour information is produced without the use of bounding boxes.
6. The system of claim 1 wherein the contour information enables an autonomous control subsystem to control a vehicle without a driver.
7. A method comprising:
receiving an input image;
producing a feature map from the input image;
applying a range of different dilation rates to the feature map to produce a final feature map maintaining a resolution corresponding to training labels, wherein the different dilation rates are applied to each of a plurality of convolution layers, the different dilation rate not having a common factor relationship other than the number one;
matching object shapes from the training labels to objects and object instances detected in the input image;
generating, based on the object shapes, contour information of the objects and object instances detected in the input image; and
applying the contour information onto the final feature map.
8. The method of claim 7 wherein the range of different dilation rates are applied in an encoding phase.
9. The method of claim 7 including applying convolutional operations directly on the feature map to generate a pixel-wise prediction map.
10. The method of claim 7 including using dense upsampling convolution with semantic segmentation.
11. The method of claim 7 wherein applying the range of different dilation rates includes using hybrid dilation convolution with semantic segmentation.
12. The method of claim 7 wherein the contour information is used by an autonomous vehicle motion planner to control a vehicle without a driver.
13. A non-transitory machine-usable storage medium embodying instructions which, when executed by at least one processor, cause the at least one processor to at least:
receive an input image;
produce a feature map from the input image;
apply a range of different dilation rates to the feature map to produce a final feature map maintaining a resolution corresponding to training labels, wherein the different dilation rates are applied to each of a plurality of convolution layers, the different dilation rate not having a common factor relationship other than the number one;
match object shapes from the training labels to objects and object instances detected in the input image;
generate, based on the object shapes, contour information of the objects and object instances detected in the input image; and
apply the contour information onto the final feature map.
14. The non-transitory machine-useable storage medium of claim 13 being configured to apply the range of different dilation rates as part of a convolution operation.
15. The non-transitory machine-useable storage medium of claim 13 wherein semantic segmentation is performed by a deep convolutional neural network trained on a cityscape dataset.
16. The non-transitory machine-useable storage medium of claim 13 being configured to use conditional random fields.
17. The non-transitory machine-useable storage medium of claim 13 wherein the contour information is produced in addition to the use of bounding boxes.
18. The non-transitory machine-useable storage medium of claim 13 wherein the contour information is used by an autonomous vehicle motion planner to plan a route for an autonomous vehicle.
19. The system of claim 1 being further configured to use dense upsampling convolution (DUC) to generate pixel-level predictions of objects detected in the input image.
20. The method of claim 7 including using dense upsampling convolution to generate pixel-level predictions of objects detected in the input image.