Systems and methods for semantically classifying and normalizing shots in video
View Patent ↗The present disclosure relates to systems and methods for classifying videos based on video content. For a given video file including a plurality of frames, a subset of frames is extracted for processing. Frames that are too dark, blurry, or otherwise poor classification candidates are discarded from the subset. Generally, material classification scores that describe type of material content likely included in each frame are calculated for the remaining frames in the subset. The material classification scores are used to generate material arrangement vectors that represent the spatial arrangement of material content in each frame. The material arrangement vectors are subsequently classified to generate a scene classification score vector for each frame. The scene classification results are averaged (or otherwise processed) across all frames in the subset to associate the video file with one or more predefined scene categories related to overall types of scene content of the video file.
1. A method for classifying a video file according to one or more scene classes, the video file including a plurality of frames, where each frame of the plurality of frames includes a plurality of pixels, and where each pixel of the plurality of pixels is associated with a vector of material classification scores describing material content in its respective frame, comprising:
for each frame of the plurality of frames, generating one or more scene classification scores associated with each of the one or more scene classes by:
dividing the frame into a plurality of grid cells;
retrieving the vector of material classification scores for each pixel in the frame;
for each grid cell of the plurality of grid cells, averaging the material classification scores across each pixel in the grid cell to form a material occurrence vector for the grid cell;
concatenating the material occurrence vectors for each grid cell of the plurality of grid cells to generate a material arrangement vector for the frame; and
based on the material arrangement vector generated for the frame, generating the one or more scene classification scores associated with each of the one or more scene classes using one or more scene classifiers;
based on the one or more scene classification scores generated for each frame of the plurality of frames, generating a representative scene classification score for each of the one or more scene classes; and
for each of the generated representative scene classification scores that is above a predetermined threshold value, labeling the video file according to the respective scene classes associated with the scene classification scores that are above the predetermined threshold value.
2. The method of claim 1 , where generating a scene classification score associated with each of the one or more scene classes is performed two or more times using varying numbers of grid cells.
3. The method of claim 2 , where the one or more scene classifiers comprise a spatial pyramid of classifiers, and where varying weights are associated with the scene classification scores.
4. The method of claim 1 , where the material content is selected from the group comprising: building, grass, person, road/sidewalk, rock, sand/gravel/soil, sky/clouds, snow/ice, trees/plants, vehicle, water, and miscellaneous.
5. The method of claim 1 , where the material occurrence vector for the grid cell represents the proportion of each type of material content included in the cell.
6. The method of claim 1 , where the material occurrence vectors for each grid cell of the plurality of grid cells are concatenated in an ordered manner.
7. The method of claim 1 , where each vector of material classification scores represents a probability that the frame includes content associated with each of the respective types of material content.
8. The method of claim 1 , where each of the one or more scene classification scores represents the probability that the frame includes content associated with each of the respective one or more scene classes.
9. The method of claim 1 , where the material arrangement vector generated for the frame represents a spatial arrangement of material content in the frame.
10. The method of claim 1 , where the one or more scene classes are selected from the group comprising: coast/beach, desert, forest, grassland, highway, indoor, lake/river, mountainous, open water, outdoor, sky, snow, and urban.
11. The method of claim 1 , where the representative scene classification score for each of the one or more scene classes comprises a statistical property of the one or more scene classification scores.
12. The method of claim 11 , where the statistical property is selected from the group comprising: average, median, maximum, and minimum.
13. The method of claim 1 , where the video file comprises a shot of video.
14. The method of claim 1 , further comprising generating a report based on the labeled video file.
15. The method of claim 1 , where one or more scene classes are used for indexing and retrieval of the video file.
16. One or more non-transitory storage media storing instructions which, when executed by one or more computing devices, cause:
classifying a video file according to one or more scene classes, the video file including a plurality of frames, where each frame of the plurality of frames includes a plurality of pixels, and where each pixel of the plurality of pixels is associated with a vector of material classification scores describing material content in its respective frame:
for each frame of the plurality of frames, generating one or more scene classification scores associated with each of the one or more scene classes by:
dividing the frame into a plurality of grid cells;
retrieving the vector of material classification scores for each pixel in the frame;
for each grid cell of the plurality of grid cells, averaging the material classification scores across each pixel in the grid cell to form a material occurrence vector for the grid cell;
concatenating the material occurrence vectors for each grid cell of the plurality of grid cells to generate a material arrangement vector for the frame; and
based on the material arrangement vector generated for the frame, generating the one or more scene classification scores associated with each of the one or more scene classes for using one or more scene classifiers;
based on the one or more scene classification scores generated for each frame of the plurality of frames, generating a representative scene classification score for each of the one or more scene classes; and
for each of the generated representative scene classification scores that is above a predetermined threshold value, labeling the video file according to the respective scene classes associated with the scene classification scores that are above the predetermined threshold value.
17. The one or more non-transitory storage media of claim 16 , where generating a scene classification score associated with each of the one or more scene classes is performed two or more times using varying numbers of grid cells.
18. The one or more non-transitory storage media of claim 16 , where the material content is selected from the group comprising: building, grass, person, road/sidewalk, rock, sand/gravel/soil, sky/clouds, snow/ice, trees/plants, vehicle, water, and miscellaneous.
19. The one or more non-transitory storage media of claim 16 , where the material occurrence vector for the grid cell represents the proportion of each type of material content included in the cell.
20. The one or more non-transitory storage media of claim 16 , where each vector of material classification scores represents a probability that the frame includes content associated with each of the respective types of material content.
21. The one or more non-transitory storage media of claim 16 , where each of the one or more scene classification scores represents the probability that the frame includes content associated with each of the respective one or more scene classes.
22. The one or more non-transitory storage media of claim 16 , where the material arrangement vector generated for the frame represents a spatial arrangement of material content in the frame.
23. The one or more non-transitory storage media of claim 16 , where the one or more scene classes are selected from the group comprising: coast/beach, desert, forest, grassland, highway, indoor, lake/river, mountainous, open water, outdoor, sky, snow, and urban.
24. The one or more non-transitory storage media of claim 16 , where the representative scene classification score for each of the one or more scene classes comprises a statistical property of the one or more scene classification scores.
25. An apparatus comprising:
a subsystem, implemented at least partially in hardware, that classifies a video file according to one or more scene classes, the video file including a plurality of frames, where each frame of the plurality of frames includes a plurality of pixels, and where each pixel of the plurality of pixels is associated with a vector of material classification scores describing material content in its respective frame:
a subsystem, implemented at least partially in hardware, that, for each frame of the plurality of frames, generates one or more scene classification scores associated with each of the one or more scene classes by:
dividing the frame into a plurality of grid cells;
retrieving the vector of material classification scores for each pixel in the frame;
for each grid cell of the plurality of grid cells, averaging the material classification scores across each pixel in the grid cell to form a material occurrence vector for the grid cell;
concatenating the material occurrence vectors for each grid cell of the plurality of grid cells to generate a material arrangement vector for the frame; and
based on the material arrangement vector generated for the frame, generating the one or more scene classification scores associated with each of the one or more scene classes for using one or more scene classifiers;
a subsystem, implemented at least partially in hardware, that, based on the one or more scene classification scores generated for each frame of the plurality of frames, generates a representative scene classification score for each of the one or more scene classes; and
a subsystem, implemented at least partially in hardware, that, for each of the generated representative scene classification scores that is above a predetermined threshold value, labels the video file according to the respective scene classes associated with the scene classification scores that are above the predetermined threshold value.