Saliency-based input resampling for efficient object detection
A processor-implemented method of video processing using includes receiving, via an artificial neural network (ANN), a video including a first frame and a second frame. A saliency map is generated based on the first frame of the video. The second frame of the video is sampled based on the saliency map. A first portion of the second frame is sampled at a first resolution and a second portion of the second frame is sampled at a second resolution. The first resolution is different than the second resolution. A resampled second frame is generated based on the sampling of the second frame. The resampled second frame is processed to determine an inference associated with the video.
1. A processor-implemented method of video processing using an artificial neural network (ANN), comprising
receiving a video including a first frame and a second frame;
processing the first frame with a first neural network model;
generating a saliency map based on the first frame of the video;
sampling the second frame of the video based on the saliency map, a first portion of the second frame being sampled at a first resolution and a second portion of the second frame being sampled at a second resolution, the first resolution being different than the second resolution;
generating a resampled second frame based on the sampling; and
processing the resampled second frame with a second neural network model to determine an inference associated with the video, wherein the second neural network model is configured with less processing capacity than the first neural network model.
2. The processor-implemented method of claim 1 , in which the first portion corresponds to locations of one of more bounding boxes included in the saliency map.
3. The processor-implemented method of claim 2 , in which the first resolution for the first portion is greater than the second resolution.
4. The processor-implemented method of claim 1 , further comprising:
determining displacements for each of a set of control points based on the saliency map;
transforming a uniform set of grid points to generate a sampling grid based on the displacements for the set of control points; and
generating the resampled second frame based on the sampling grid.
5. The processor-implemented method of claim 4 , in which the resampled second frame is generated based on a learned zooming behavior.
6. The processor-implemented method of claim 1 , further comprising sampling a third portion of the second frame at a third resolution based on the saliency map, the third resolution being different than the first resolution and the second resolution.
7. An apparatus for video processing using an artificial neural network (ANN), comprising:
a memory; and
at least one processor coupled to the memory, the at least one processor configured to:
receive a video including a first frame and a second frame;
process the first frame with a first neural network model;
generate a saliency map based on the first frame of the video;
sample the second frame of the video based on the saliency map, a first portion of the second frame being sampled at a first resolution and a second portion of the second frame being sampled at a second resolution, the first resolution being different than the second resolution;
generate a resampled second frame based on the sampling; and
process the resampled second frame with a second neural network model to determine an inference associated with the video, wherein the second neural network model is configured with less processing capacity than the first neural network model.
8. The apparatus of claim 7 , in which the first portion corresponds to locations of one of more bounding boxes included in the saliency map.
9. The apparatus of claim 8 , in which the first resolution for the first portion is greater than the second resolution.
10. The apparatus of claim 7 , in which the at least one processor is further configured to:
determine displacements for each of a set of control points based on the saliency map;
transform a uniform set of grid points to generate a sampling grid based on the displacements for the set of control points; and
generate the resampled second frame based on the sampling grid.
11. The apparatus of claim 10 , in which the at least one processor is further configured to generate the resampled second frame based on a learned zooming behavior.
12. The apparatus of claim 7 , in which the at least one processor is further configured to sample a third portion of the second frame at a third resolution based on the saliency map, the third resolution being different than the first resolution and the second resolution.
13. A non-transitory computer-readable medium having program code recorded thereon for video processing using an artificial neural network (ANN), the program code being executed by a processor and comprising:
program code to receive a video including a first frame and a second frame;
program code to process the first frame with a first neural network model;
program code to generate a saliency map based on the first frame of the video;
program code to sample the second frame of the video based on the saliency map, a first portion of the second frame being sampled at a first resolution and a second portion of the second frame being sampled at a second resolution, the first resolution being different than the second resolution;
program code to generate a resampled second frame based on the sampling; and
program code to process the resampled second frame with a second neural network model to determine an inference associated with the video, wherein the second neural network model is configured with less processing capacity than the first neural network model.
14. The non-transitory computer-readable medium of claim 13 , in which the first portion corresponds to locations of one of more bounding boxes included in the saliency map.
15. The non-transitory computer-readable medium of claim 14 , in which the first resolution for the first portion is greater than the second resolution.
16. The non-transitory computer-readable medium of claim 13 , further comprising:
program code to determine displacements for each of a set of control points based on the saliency map;
program code to transform a uniform set of grid points to generate a sampling grid based on the displacements for the set of control points; and
program code to generate the resampled second frame based on the sampling grid.
17. The non-transitory computer-readable medium of claim 16 , further comprising program code to generate the resampled second frame based on a learned zooming behavior.
18. The non-transitory computer-readable medium of claim 13 , further comprising program code to sample a third portion of the second frame at a third resolution based on the saliency map, the third resolution being different than the first resolution and the second resolution.