Loading patent…
AssigneeOracle International Corporation
InventorsAnkit Kumar Aggarwal, Reetesh Mukul, Pramir Sarkar
Estimating the term…
Disclosure relates to training a model to score video clips based on input prompts and associated ground truth information. Training datapoints may include a training video clip, an input prompt, and ground truth information. The model may generate a score for each clip based at least in part on the input prompt. A loss calculation subsystem may compute a loss using the score and the ground truth information. A loss minimization subsystem may update model parameters to train the model. In some embodiments, similarity scores may be generated for frames or images of a training video clip and/or be adjusted using a Weibull-based methodology to produce an overall clip score. In some embodiments, training may use an N-pair contrastive loss such that a clip annotated with a central time may receive a highest score while boundary clips annotated with a start time and/or end time retain partial similarity.
A method, comprising:
The method of claim , wherein generating the score for the training video clip comprises generating a similarity score for each frame of a set of frames corresponding to the training video clip.
The method of claim , wherein generating the score for the training video clip comprises using a Weibull-based methodology to adjust distribution of the set of frames and associated scores within the training video clip to generate an overall score of the training video clip.
The method of claim , wherein the ground truth information comprises a target score for the training video clip using the Weibull-based methodology.
The method of claim , wherein the input prompt comprises one or more user queries including criteria and preferences.
The method of claim , wherein the training video clip comprises a segment of a sporting event, and wherein the ground truth information identifies an occurrence using a tuple comprising at least one of a start-time, a central time, or an end-time associated with the sporting event.
The method of claim , wherein the loss minimization comprises a calculation utilizing a modified N-pair loss of contrastive learning.
The method of claim , wherein the modified N-pair loss of contrastive learning enables a certain training video clip containing the central time to have a highest score.
The method of claim , wherein the modified N-pair loss of contrastive learning enables a second training video clip including at least one of the start-time or the end-time to have a certain score that indicates partial similarity.
The method of claim , wherein performing the loss minimization comprises updating the one or more model parameters using backward propagation.
The method of claim , wherein the machine learning (ML) model comprises a highlight clip estimator.
The method of claim , wherein a training environment further comprises an optimizer, and wherein the ground truth information comprises a target video highlight summary.
The method of claim , further comprising:
A nontransitory computer-readable medium storing computer-executable instructions that, when executed by one or more processors of a computing system, cause the computing system to perform operations comprising:
The nontransitory computer-readable medium of claim wherein the ground truth information identifies an event using a tuple comprising at least one of a start-time, a central time, and an end-time associated with the event.
The nontransitory computer-readable medium of claim wherein the loss minimization comprises a calculation utilizing a modified N-pair loss of contrastive learning.
The nontransitory computer-readable medium of claim wherein the modified N-pair loss of contrastive learning enables a certain training video clip containing the central time to have a highest score.
The nontransitory computer-readable medium of claim wherein the modified N-pair loss of contrastive learning enables a second training video clip including at least one of the start-time or the end-time to have a certain score that indicates partial similarity.
A computing system, comprising:
The computing system of claim wherein generating the score for the training video clip comprises:
Dependencies are inferred from the claim text.