Training data generation for advanced frequency management
Systems and methods for programmatic generation of training data, including: a training data generation engine configured to: identify an image asset corresponding to an entity; identify a training video; select a consecutive subset of frames of the training video based on a procedure for ranking frames on their candidacy for overlaying content; for at least one frame of the subset of frames: perform an augmentation technique on the identified logo image to generate an augmented image asset; overlay at least one variation of the image asset, including the augmented image asset, onto each of the subset of frames to generate a set of overlayed frames; and generate an augmented version of the training video including the overlayed frames; and a model training engine configured to: train an artificial intelligence model for entity detection using the augmented version of the training video.
1. A system for programmatic generation of training data, comprising:
a computer processor; and
a training data generation engine executing on the computer processor and configured to enable the computer processor to:
identify an image asset corresponding to an entity;
identify a training video;
select a consecutive subset of frames of the training video based on a procedure for ranking frames on their candidacy for overlaying content, wherein the procedure evaluates presence of surfaces suitable for content overlay;
for at least one frame of the subset of frames: perform an augmentation technique on the identified logo image to generate an augmented image asset;
overlay at least one variation of the image asset, including the augmented image asset, onto each of the subset of frames to generate a set of overlayed frames; and
generate an augmented version of the training video comprising the overlayed frames; and
a model training engine configured to:
train an artificial intelligence model for entity detection using the augmented version of the training video; and
a deep learning model service configured to execute the trained artificial intelligence model on a set of video advertisements to identify brand identifiers associated with a set of entities; and
an offline transcoding service configured to store the brand identifiers associated with the set of entities in a repository; and
an online media service configured to:
identify a set of frequency thresholds associated with the brand identifiers;
calculate frequency metrics of the brand identifiers based on frequency of serving media content associated with the brand identifiers to user clients; and
regulate serving of media content to user clients to avoid exceeding the frequency thresholds based on the calculated frequency metrics.
2. The system of claim 1 , wherein the procedure for ranking frames on their candidacy for overlaying content comprises:
using an object detection algorithm to identify a set of consecutive frames each containing an area suitable for brand placement;
determining that the set of consecutive frames meet a predefined set of criteria for overlaying content; and
calculating a ranking score for the set of consecutive frames, the ranking score indicating the suitability for overlaying at least one variation of the image asset on the detected area within each frame.
3. The system of claim 2 , wherein the predefined set of criteria comprises a maximum movement threshold indicating an amount of spatial variation of the area within each consecutive frame of the set of consecutive frames.
4. The system of claim 3 , wherein overlaying at least one variation of the image asset onto each of the subset of frames creates the appearance of movement of the identified logo image in the set of consecutive frames.
5. The system of claim 2 , wherein the predefined set of criteria comprises a minimum cardinality of the set of consecutive frames.
6. The system of claim 1 , wherein the training video is relevant to a domain, and wherein the entity is determined to be relevant to the same domain.
7. The system of claim 1 , wherein the training data generation engine performing the augmentation technique comprises:
performing a transformation operation on the identified logo image, wherein the transformation operation is at least one selected from the group consisting of a skew operation, a scale operation, a translation operation, and a rotation operation, and wherein the transformation operation is at least partly based on an area in each of the subset of frames upon which the augmented image asset is to be overlayed.
8. The system of claim 1 , wherein the model training engine is further configured to:
use gradient descent in order to tune parameters of the model to maximize the fit of a set of prediction data to a ground truth dataset; and
perform hyperparameter tuning to exclude false positives and to optimize a precision value and a recall value associated with the model.
9. The system of claim 8 , wherein the training data generation engine performing hyperparameter tuning comprises:
executing the model on a validation dataset to identify a set of candidate entity/probability pairs; and
executing a voting algorithm using a candidate threshold and a score threshold to exclude at least one of the candidate entity/probability pairs from a final result set, wherein the candidate threshold value represents a minimum number of valid detections within the candidate pairs in order for an entity to be included in the result set, and wherein the score threshold represents a minimum confidence score required to consider detection of an entity valid.
10. A method for programmatic generation of training data, comprising:
identifying an image asset corresponding to an entity;
identifying a training video;
selecting a consecutive subset of frames of the training video based on a procedure for ranking frames on their candidacy for overlaying content, wherein the procedure evaluates presence of surfaces suitable for content overlay;
for at least one frame of the subset of frames: performing, using at least one computer processor, an augmentation technique on the identified logo image to generate an augmented image asset;
overlaying at least one variation of the image asset, including the augmented image asset, onto each of the subset of frames to generate a set of overlayed frames; and
generating an augmented version of the training video comprising the overlayed frames; and
training an artificial intelligence model for entity detection using the augmented version of the training video;
executing the trained artificial intelligence model on a set of video advertisements to identify brand identifiers associated with a set of entities;
storing the brand identifiers associated with the set of entities in a repository;
identifying a set of frequency thresholds associated with the brand identifiers; and
calculating frequency metrics of the brand identifiers based on frequency of serving media content associated with the brand identifiers to user clients; and
regulating serving of media content to user clients to avoid exceeding the frequency thresholds based on the calculated frequency metrics.
11. The method of claim 10 , wherein the procedure for ranking frames on their candidacy for overlaying content comprises:
using an object detection algorithm to identify a set of consecutive frames each containing an area suitable for brand placement;
determining that the set of consecutive frames meet a predefined set of criteria for overlaying content; and
calculating a ranking score for the set of consecutive frames, the ranking score indicating the suitability for overlaying at least one variation of the image asset on the detected area within each frame.
12. The method of claim 11 , wherein the predefined set of criteria comprises a maximum movement threshold indicating an amount of spatial variation of the area within each consecutive frame of the set of consecutive frames.
13. The method of claim 12 , wherein overlaying at least one variation of the image asset onto each of the subset of frames creates the appearance of movement of the identified logo image in the set of consecutive frames.
14. The method of claim 11 , wherein the predefined set of criteria comprises a minimum cardinality of the set of consecutive frames.
15. The method of claim 10 , wherein the augmentation technique comprises:
performing a transformation operation on the identified logo image, wherein the transformation operation is at least one selected from the group consisting of a skew operation, a scale operation, a translation operation, and a rotation operation, and wherein the transformation operation is at least partly based on an area in each of the subset of frames upon which the augmented image asset is to be overlayed.
16. The method of claim 10 , further comprising:
using gradient descent in order to tune parameters of the model to maximize the fit of a set of prediction data to a ground truth dataset; and
performing hyperparameter tuning to exclude false positives and to optimize a precision value and a recall value associated with the model.
17. The method of claim 16 , wherein hyperparameter tuning comprises:
executing the model on a validation dataset to identify a set of candidate entity/probability pairs; and
executing a voting algorithm using a candidate threshold and a score threshold to exclude at least one of the candidate entity/probability pairs from a final result set, wherein the candidate threshold value represents a minimum number of valid detections within the candidate pairs in order for an entity to be included in the result set, and wherein the score threshold represents a minimum confidence score required to consider detection of an entity valid.
18. A non-transitory computer-readable storage medium comprising a plurality of instructions for programmatic generation of training data, the plurality of instructions configured to execute on at least one computer processor to enable the at least one computer processor to:
identify an image asset corresponding to an entity;
identify a training video;
select a consecutive subset of frames of the training video based on a procedure for ranking frames on their candidacy for overlaying content, wherein the procedure evaluates presence of surfaces suitable for content overlay;
for at least one frame of the subset of frames: perform an augmentation technique on the identified logo image to generate an augmented image asset;
overlay at least one variation of the image asset, including the augmented image asset, onto each of the subset of frames to generate a set of overlayed frames; and
generate an augmented version of the training video comprising the overlayed frames; and
train an artificial intelligence model for entity detection using the augmented version of the training video;
execute the trained artificial intelligence model on a set of video advertisements to identify brand identifiers associated with a set of entities;
store the brand identifiers associated with the set of entities in a repository;
identify a set of frequency thresholds associated with the brand identifiers; and
calculate frequency metrics of the brand identifiers based on frequency of serving media content associated with the brand identifiers to user clients, wherein the frequency metrics are configured for regulating serving of media content to user clients to avoid exceeding the frequency thresholds.