Fixation generation for machine learning
The disclosure extends to methods, systems, and apparatuses for automated fixation generation and more particularly relates to generation of synthetic saliency maps. A method for generating saliency information includes receiving a first image and an indication of one or more sub-regions within the first image corresponding to one or more objects of interest. The method includes generating and storing a label image by creating an intermediate image having one or more random points. The random points have a first color in regions corresponding to the sub-regions and a remainder of the intermediate image having a second color. Generating and storing the label image further includes applying a Gaussian blur to the intermediate image.
1. A method comprising:
identifying a ground truth bounding box corresponding to an object of interest in a first image;
executing a randomization algorithm to generate an intermediate image comprising one or more random points within a region corresponding to the ground truth bounding box of the first image, wherein a quantity of random points is determined based on a size of the ground truth bounding box;
generating a blurred intermediate image by applying a blur to each of the one or more random points in the intermediate image; and
storing the blurred intermediate image as a label image for the first image.
2. The method of claim 1 , further comprising pairing the first image and the label image and feeding the pair of images to a modified deep neural network configured to output a synthetic saliency map to predict a location of an object within any image.
3. The method of claim 1 , wherein the blur applied to each of the one or more random points in the intermediate image comprises an ellipses shaped blur configured to predict a scale and location of the object of interest.
4. The method of claim 1 , further comprising downsizing the blurred intermediate image to generate a low-resolution version of the blurred intermediate image, and wherein storing the blurred intermediate image as the label image comprises storing the low-resolution version of the blurred intermediate image.
5. The method of claim 1 , further comprising training a neural network to determine information about any object of interest in an image by using the first image and the label image as training data for the neural network.
6. The method of claim 5 , wherein the neural network consumes the first image and the label image and outputs a synthetic saliency map based on the first image and label image.
7. The method of claim 6 , wherein the synthetic saliency map mimics human perception for object detection.
8. The method of claim 1 , further comprising calculating the quantity of random points to generate within the ground truth bounding box based on the size of the ground truth bounding box.
9. The method of claim 1 , wherein generating the blurred intermediate image comprises applying a Gaussian blur to the intermediate image.
10. A system comprising one or more processors that are programmable to execute instructions stored in non-transitory computer readable storage media, the instructions comprising
identifying a ground truth bounding box corresponding to an object of interest in a first image;
applying a randomization algorithm to generate an intermediate image comprising one or more random points within a region corresponding to the ground truth bounding box of the first image, wherein a quantity of random points is determined based on a size of the ground truth bounding box;
generating a blurred intermediate image by applying a blur to each of the one or more random points in the intermediate image; and
storing the blurred intermediate image as a label image for the first image.
11. The system of claim 10 , wherein the instructions further comprise pairing the first image and the label image and feeding the pair of image to a modified deep neural network configured to output a synthetic saliency map to predict a location of an object within any image.
12. The system of claim 10 , wherein the instructions further comprise receiving ground truth about the object of interest within the ground truth bounding box, wherein the ground truth comprises one or more of a classification, an orientation, and a relative location of the object of interest.
13. The system of claim 10 , wherein a machine learning algorithm or model is configured to consume the first image and the label image and output a synthetic saliency map, wherein the synthetic saliency map mimics human perception for object detection.
14. The system of claim 10 , wherein the blur applied to each of the one or more random points in the intermediate image comprises an ellipses shaped blur configured to predict a scale and location of the object of interest.
15. The system of claim 10 , wherein the instructions further comprise calculating the quantity of random points to generate within the ground truth bounding box based on the size of the ground truth bounding box.
16. Non-transitory computer readable storage media storing instructions to be executed by one or more processors, the instructions comprising:
identifying a ground truth bounding box corresponding to an object of interest in a first image;
applying a randomization algorithm to generate an intermediate image comprising one or more random points within a region corresponding to the ground truth bounding box of the first image, wherein a quantity of random points is determined based on a size of the ground truth bounding box;
generating a blurred intermediate image by applying a blur to each of the one or more random points in the intermediate image; and
storing the blurred intermediate image as a label image for the first image.
17. The non-transitory computer readable storage media of claim 16 , wherein the instructions further comprise pairing the first image and the label image and feeding the pair of images to a modified deep neural network configured to output a synthetic saliency map to predict a location of an object within any image.
18. The non-transitory computer readable storage media of claim 16 , wherein the instructions are such that blurring each of the one or more random points in the intermediate image comprises applying a Gaussian blur to generate an ellipses shaped blur configured to predict a scale and location of the object of interest.
19. The non-transitory computer readable storage media of claim 16 , wherein the instructions further comprise training a neural network to generate a synthetic saliency map that mimics human perception for object detection by providing the blurred intermediate image and the label image to the neural network as a training dataset.
20. The non-transitory computer readable storage media of claim 16 , wherein the instructions further comprise calculating the quantity of random points to generate within the ground truth bounding box based on the size of the ground truth bounding box.