Object detection improvement using a foreground occupancy map
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for obtaining a foreground occupancy map for a camera view. The methods, systems, and apparatus include actions of determining an area of an image in which there is a false detection of an object, determining a likely contribution of the area to the false detection based on the foreground occupancy map, generating a modified object detector based on the likely contribution of the area, and detecting an object using the modified object detector.
1. A computer-implemented method comprising:
obtaining a foreground occupancy map for a camera view;
determining an area of an image in which there is a false detection of an object;
determining a likely contribution of the area to the false detection based on the foreground occupancy map;
generating a modified object detector based on the likely contribution of the area comprising:
increasing a loss component for a bounding box based on the likely contribution of the area; and
training the modified object detector based on the loss component; and
detecting an object using the modified object detector.
2. The method of claim 1 , wherein obtaining a foreground occupancy map for a camera view comprises:
determining a frequency that pixels within images from the camera view are included in a bounding box for images in a training dataset.
3. The method of claim 2 , wherein each of the images in the training dataset includes a respective bounding box.
4. The method of claim 2 , wherein the bounding box indicates that pixels within the bounding box include an object of interest.
5. The method of claim 1 , wherein determining an area of an image in which there is a false detection of an object comprises:
determining that a bounding box generated for the image does not include an object of interest.
6. The method of claim 5 , wherein determining that a bounding box generated for the image does not include an object of interest comprises:
providing the image to an object detector generated from images in a training dataset; and
receiving, from the object detector, an indication of the bounding box.
7. The method of claim 1 , wherein determining a likely contribution of the area to the false detection based on the foreground occupancy map comprises:
determining values for pixels in the foreground occupancy map for the camera view that correspond to pixels in the bounding box; and
determining the likely contribution of the area to the false detection based on the values for the pixels in the foreground occupancy map.
8. A system comprising:
one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
obtaining a foreground occupancy map for a camera view;
determining an area of an image in which there is a false detection of an object;
determining a likely contribution of the area to the false detection based on the foreground occupancy map;
generating a modified object detector based on the likely contribution of the area, comprising:
increasing a loss component for a bounding box based on the likely contribution of the area; and
training the modified object detector based on the loss component; and
detecting an object using the modified object detector.
9. The system of claim 8 , wherein obtaining a foreground occupancy map for a camera view comprises:
determining a frequency that pixels within images from the camera view are included in a bounding box for images in a training dataset.
10. The system of claim 9 , wherein each of the images in the training dataset includes a respective bounding box.
11. The system of claim 9 , wherein the bounding box indicates that pixels within the bounding box include an object of interest.
12. The system of claim 8 , wherein determining an area of an image in which there is a false detection of an object comprises:
determining that a bounding box generated for the image does not include an object of interest.
13. The system of claim 12 , wherein determining that a bounding box generated for the image does not include an object of interest comprises:
providing the image to an object detector generated from images in a training dataset; and
receiving, from the object detector, an indication of the bounding box.
14. The system of claim 8 , wherein determining a likely contribution of the area to the false detection based on the foreground occupancy map comprises:
determining values for pixels in the foreground occupancy map for the camera view that correspond to pixels in the bounding box; and
determining the likely contribution of the area to the false detection based on the values for the pixels in the foreground occupancy map.
15. A non-transitory computer-readable medium storing software comprising instructions executable by one or more computers which, upon such execution, cause the one or more computers to perform operations comprising:
obtaining a foreground occupancy map for a camera view;
determining an area of an image in which there is a false detection of an object;
determining a likely contribution of the area to the false detection based on the foreground occupancy map;
generating a modified object detector based on the likely contribution of the area comprising:
increasing a loss component for a bounding box based on the likely contribution of the area; and
training the modified object detector based on the loss component; and
detecting an object using the modified object detector.
16. The medium of claim 15 , wherein obtaining a foreground occupancy map for a camera view comprises:
determining a frequency that pixels within images from the camera view are included in a bounding box for images in a training dataset.
17. The medium of claim 16 , wherein each of the images in the training dataset includes a respective bounding box.
18. The medium of claim 16 , wherein the bounding box indicates that pixels within the bounding box include an object of interest.
19. The medium of claim 15 , wherein determining an area of an image in which there is a false detection of an object comprises:
determining that a bounding box generated for the image does not include an object of interest.
20. The medium of claim 19 , wherein determining that a bounding box generated for the image does not include an object of interest comprises:
providing the image to an object detector generated from images in a training dataset; and
receiving, from the object detector, an indication of the bounding box.