Generating a highlight video from an input video

1. A computer implemented method of generating at least one highlight video from an input video, comprising, using at least one processor for:

identifying a plurality of significant frames of the input video;

computing video-level features of the input video;

selecting a plurality of subsets of the plurality of significant frames;

for each subset, computing a similarity score indicating similarity between visual features of the subset and video-level features of the input video;

clustering the input video into a plurality of clusters of sequential frames according to sequential positions within the video based on the similarity scores correlated with the plurality of significant frames;

creating at least one highlight video by selecting a cluster of sequential frames of the input video;

computing a normalized score by normalizing the similarity scores of the at least one highlight video, and presenting each of the at least one highlight video a relative number of times corresponding to the normalized similarity score, wherein the at least one highlight video are presented to a plurality of client terminals accessing a web page;

testing the at least one highlight video presented the relative number of times corresponding to the normalized similarity score to obtain a click rate for each of the at least one highlight video;

generating a multi-record training dataset, wherein a record comprises a sample main video, and a ground truth of the at least one highlight video generated from the sample main video each labelled with a corresponding click rate; and

training a machine learning model for generating an outcome of a target highlight video predicted to generate a highest click rate in response to an input of a target main video.

2. The computer implemented method of claim 1 , further comprising:

providing the at least one highlight video for playing on a webpage by a video player without audio.

3. The computer implemented method of claim 1 , further comprising automatically generating a hyperlink to the input video from the at least one highlight video.

4. The computer implemented method of claim 1 , wherein clustering comprises:

defining a dataset of sequential positions of frames of the video mapped to the similarity scores according to sequential position of the significant frames of each subset associated with each similarity score;

identifying a plurality of maximum values within the dataset; and

clustering the frames of the video according to each maximum value.

5. The computer implemented method of claim 4 , further comprising applying a kernel function to the dataset to obtained a smoothed dataset, wherein the plurality of maximum values are identified in the smoothed dataset.

6. The computer implemented method of claim 1 , further comprising:

defining a sliding window having a duration of the highlight video;

iteratively applying the sliding window to the plurality of significant frames, wherein the subset includes significant frames falling within a current application of the sliding window.

7. The computer implemented method of claim 6 , wherein the sliding window is iteratively slide to a next significant frame.

8. The computer implemented method of claim 1 , wherein the similarity score is computed according to at least one value associated with each category identified for the subset that match at least one value associated with categories of the video-level features, for a number of significant frames of the subset.

9. The computer implemented method of claim 1 , wherein the visual features of the subset exclude audio of the input video and the video-level features of the input video is computed based on audio of the input video.

10. The computer implemented method of claim 1 , wherein creating the at least one highlight video comprises creating a plurality of highlight videos according to a decreased ranking of clusters from a highest similarity score.

11. The computer implemented method of claim 1 , wherein creating comprises generating instructions for selectively playing the selected cluster of sequential frames of the input video as the at least one highlight video without creating a new video file.

12. The computer implemented method of claim 1 , wherein creating comprises selecting the sequential frames of the selected cluster as the highlight video.

13. The computer implemented method of claim 1 , further comprising:

computing a probability score by normalizing the similarity scores of the at least one highlight video;

creating a multi-record training dataset, wherein a record comprises: a sample main video, and a ground truth of the at least one highlight video generated from the sample main video each labelled by the corresponding probability score; and

training a machine learning model for generating an outcome of a target highlight video representing a highest similarity to a target main video in response to an input of the target main video.

14. The computer implemented method of claim 1 , further comprising:

creating a multi-record training dataset, wherein a record comprises visual features of the subset, video-level features of the input video, and a ground truth label of the computed similarity score;

training a similarity machine learning model; and

computing the similarity score for a new input video by feeding the visual features of the subset and video-level feature of the new input video into the similarity machine learning model.

15. The computer implemented method of claim 1 , further comprising: selecting a portion of the cluster having highest similarity score, the portion of the cluster having a time duration of the highlight video.

16. The computer implemented method of claim 1 , wherein a current frame is selected as a significant frame when change information indicative of content of the current frame is sufficiently different than a preceding significant frame.

17. The computer implemented method of claim 1 , wherein the plurality of significant frames are selected by:

obtaining change information of the input video constructed of a plurality of encoded frames, the change information created for each of the plurality of encoded frames by an encoder is indicative of a difference between a visual content of a respective encoded frame and visual content of a predecessor encoded frame;

performing following actions for each of the plurality of encoded frames:

analyze the change information to calculate a cumulative difference between visual content of the respective encoded frame and visual content of a most recently selected significant frame previously selected from the plurality of encoded frames, and

select the respective encoded frame as another significant frame in case the cumulative difference exceeds a predefined threshold.

18. The computer implemented method of claim 1 , wherein the visual features of the subset are computed as an aggregation of visual features of each significant frame of the subset, wherein visual features of each significant frame is computed by recognizing at least one object or text in the significant frame.

19. The computer implemented method of claim 1 , wherein the video-level features of the video is computed as a combination of an aggregation of visual features of the plurality of significant frames of the input video and an analysis of the input video as a whole.

20. The computer implemented method of claim 19 , wherein the analysis of the input video as a whole comprises extracting content features from the video using at least one approach from a group comprising: speech converted to text, title, description, metadata, natural language processing of textual data, optical character recognition, list of people in the input video, keywords mapped to the input video, categories assigned to the input video, and objects identified in the input video.

21. A system for generating at least one highlight video from an input video, comprising:

at least one processor executing a code for:

identifying a plurality of significant frames of the input video;

computing video-level features of the input video;

selecting a plurality of subsets of the plurality of significant frames;

for each subset, computing a similarity score indicating similarity between visual features of the subset and video-level features of the input video;

clustering the input video into a plurality of clusters of sequential frames according to sequential positions within the video based on the similarity scores correlated with the plurality of significant frames; and

creating at least one highlight video by selecting a cluster of sequential frames of the input video;

computing a normalized score by normalizing the similarity scores of the at least one highlight video, and presenting each of the at least one highlight video a relative number of times corresponding to the normalized similarity score, wherein the at least one highlight video are presented to a plurality of client terminals accessing a web page;

testing the at least one highlight video presented the relative number of times corresponding to the normalized similarity score to obtain a click rate for each of the at least one highlight video;

generating a multi-record training dataset, wherein a record comprises a sample main video, and a ground truth of the at least one highlight video generated from the sample main video each labelled with a corresponding click rate; and

training a machine learning model for generating an outcome of a target highlight video predicted to generate a highest click rate in response to an input of a target main video.

22. A non-transitory medium storing program instructions for generating at least one highlight video from an input video, which, when executed by at least one processor, cause the at least one processor to:

identify a plurality of significant frames of the input video;

compute video-level features of the input video;

select a plurality of subsets of the plurality of significant frames;

for each subset, compute a similarity score indicating similarity between visual features of the subset and video-level features of the input video;

cluster the input video into a plurality of clusters of sequential frames according to sequential positions within the video based on the similarity scores correlated with the plurality of significant frames; and

create at least one highlight video by selecting a cluster of sequential frames of the input video;

computing a normalized score by normalizing the similarity scores of the at least one highlight video, and presenting each of the at least one highlight video a relative number of times corresponding to the normalized similarity score, wherein the at least one highlight video are presented to a plurality of client terminals accessing a web page;

testing the at least one highlight video presented the relative number of times corresponding to the normalized similarity score to obtain a click rate for each of the at least one highlight video;

generating a multi-record training dataset, wherein a record comprises a sample main video, and a ground truth of the at least one highlight video generated from the sample main video each labelled with a corresponding click rate; and

training a machine learning model for generating an outcome of a target highlight video predicted to generate a highest click rate in response to an input of a target main video.