Automated generation of training data for contextually generated perceptions
Embodiments of the present invention train multiple Perception models to predict contextual metadata (tags) with respect to target content items. By extracting context from content items, and generating associations among the Perception models, individual Perceptions trigger one another based on the extracted context to generate a more robust set of contextual metadata. A Perception Identifier predicts core tags that make coarse distinctions among content items at relatively higher levels of abstraction, while also triggering other Perception models to predict additional perception tags at lower levels of abstraction. A Dense Classifier identifies sub-content items at various levels of abstraction, and facilitates the iterative generation of additional dense tags across integrated Perceptions. Class-specific thresholds are generated with respect to individual classes of each Perception to address the inherent sampling bias that results from the varying number and quality of training samples (across different classes of content items) available to train each Perception.
1. A method for automatically generating training data with respect to an object contained within a target video, the method comprising the following steps:
(a) receiving a first frame of the target video, wherein the first frame includes a bounding box containing the object and a tag associated with the object;
(b) generating a model defining the object within the bounding box, wherein the model has not yet been trained to detect the object;
(c) receiving a plurality of subsequent frames of the target video;
(d) employing the model to track the object across each of the subsequent frames; and
(e) tagging each of the subsequent frames containing the object with the tag, thereby generating a set of training data that can be employed to train the model to detect objects similar to the object when presented with other images or target videos not previously encountered by the model.
2. The method of claim 1 , further comprising the following steps:
(a) generating the model by generating segmentation data defining the object; and
(b) employing the segmentation data to track the object across the subsequent frames of the target video.
3. The method of claim 1 , wherein a plurality of sets of training data are generated automatically with respect to a respective plurality of objects and associated tags.
4. The method of claim 3 , further comprising the following steps:
(a) generating (i) a first set of training data with respect to a first object and an associated first tag and (ii) a second set of training data with respect to a second object and an associated second tag;
(b) training a model to distinguish the first tag from the second tag by receiving both the first set of training data and the second set of training data;
(c) generating a first class-specific threshold with respect to the first tag;
(d) generating a second class-specific threshold with respect to the second tag; and
(e) when the model is presented with a target image, utilizing (i) the first class-specific threshold to determine whether the model predicts the first tag and (ii) the second class-specific threshold to determine whether the model predicts the second tag.
5. The method of claim 3 , further comprising the following steps:
(a) generating (i) a first set of training data with respect to a first object and an associated first tag, (ii) a second set of training data with respect to a second object and an associated second tag, (iii) a third set of training data with respect to a third object and an associated third tag, (ii) a fourth set of training data with respect to a fourth object and an associated fourth tag;
(b) training a first model to distinguish the first tag from the second tag by receiving both the first set of training data and the second set of training data;
(c) training a second model to distinguish the third tag from the fourth tag by receiving both the third set of training data and the fourth set of training data; and
(d) generating an association between (i) a subset of tags generated by the first model when presented with the third set of training data and (ii) the third tag;
(e) wherein the first model, when presented with a target image, generates a set of predicted tags and, if the set of predicted tags matches the subset of tags in the association, triggers the second model to predict the third tag.
6. The method of claim 3 , wherein, for each of a plurality of target videos, a plurality of sets of training data are generated automatically with respect to a respective plurality of objects and associated tags.