Method for generating video- and audience-specific encoding ladders

1. A method for generating an encoding ladder comprising:

extracting a set of video features representing properties of a video segment;

generating a feature vector for the video segment based on the set of video features;

passing the feature vector into a convex hull estimation model to generate a set of bitrate-resolution pairs based on the set of video features, each bitrate-resolution pair in the set of bitrate-resolution pairs:

defining a bitrate; and

defining a resolution predicted to yield a quality score, approximating a maximum quality score, characterizing the video segment encoded at the bitrate and the resolution;

accessing a distribution of audience bandwidths representative of likely viewers of the video segment;

selecting a highest bitrate-resolution pair for the encoding ladder from the set of bitrate-resolution pairs, the highest bitrate-resolution pair defining a highest bitrate;

selecting a lowest bitrate-resolution pair for the encoding ladder from the set of bitrate-resolution pairs, the lowest bitrate-resolution pair defining a lowest bitrate;

selecting a subset of bitrate-resolution pairs from the set of bitrate-resolution pairs based on the distribution of audience bandwidths, the subset of bitrate-resolution pairs defining bitrates less than the highest bitrate and greater than the lowest bitrate; and

generating the encoding ladder, for the video segment, comprising the highest bitrate-resolution pair, the lowest bitrate-resolution pair, and the subset of bitrate-resolution pairs.

2. The method of claim 1 , further comprising:

encoding a set of renditions of the video segment, each rendition comprising an encoding of the video segment characterized by a bitrate and a resolution of a bitrate-resolution pair in the encoding ladder; and

publishing a manifest file representing the encoding ladder for an internet stream.

3. The method of claim 1 , further comprising:

accessing a set of training video segments;

for each training video segment in the set of training video segments:

extracting a set of training video features representing properties of the training video segment;

generating a training feature vector for the training video segment based on the set of training video features;

encoding a set of trial encodes of the training video segment, the trial encodes spanning a set of encoding bitrates;

calculating a quality score of each trial encode in the set of trial encodes according to a quality metric to generate a set of quality scores;

for each encoding bitrate in the set of encoding bitrates:

identifying a subset of trial encodes characterized by the encoding bitrate; and

selecting a trial encode for a convex hull of the training video from the subset of trial encodes, the trial encode characterized by a maximum quality score relative to the subset of trial encodes;

generating a training example comprising the training feature vector and the convex hull of the training video segment; and

storing the training example in a set of training examples; and

training the convex hull estimation model based on the set of training examples.

4. The method of claim 1 , wherein selecting the highest bitrate-resolution pair in the set of bitrate-resolution pairs comprises:

for each bitrate-resolution pair in the set of bitrate-resolution pairs, predicting a quality score characterizing the bitrate-resolution pair according to a quality metric;

identifying, in the set of bitrate-resolution pairs, a set of high-quality bitrate-resolution pairs characterized by quality scores greater than a threshold quality score; and

selecting the highest bitrate-resolution pair from the set of high-quality bitrate-resolution pairs, the highest bitrate-resolution pair defining a minimum bitrate in the set of high-quality bitrate resolution pairs.

5. The method of claim 1 , wherein selecting the lowest bitrate-resolution pair in the set of bitrate-resolution pairs comprises:

identifying a threshold percentile bandwidth in the distribution of audience bandwidths; and

selecting a first bitrate-resolution pair defining a bitrate greater than the threshold percentile bandwidth as the lowest bitrate-resolution pair.

6. The method of claim 1 , wherein selecting the subset of bitrate-resolution pairs in the set of bitrate-resolution pairs based on the distribution of audience bandwidths comprises selecting the subset of bitrate resolution pairs predicted to yield an aggregate quality score approximating a maximum aggregate quality score of the video segment viewed according to the distribution of audience bandwidths.

7. A method for generating an encoding ladder comprising:

accessing a set of training video segments;

for each training video segment in the set of training video segments:

extracting a set of training video features representing properties of the training video segment;

encoding a set of trial encodes of the training video segment, the trial encodes spanning a set of encoding bitrates;

calculating a quality score of each trial encode in the set of trial encodes according to a quality metric to generate a set of quality scores;

for each encoding bitrate in the set of encoding bitrates:

identifying a subset of trial encodes characterized by the encoding bitrate; and

selecting a trial encode for a convex hull of the training video from the subset of trial encodes, the trial encode characterized by a maximum quality score relative to the subset of trial encodes;

generating a training example comprising the set of training video features of the training video segment and the convex hull of the training video segment; and

storing the training example in a set of training examples;

training the convex hull estimation model based on the set of training examples;

extracting a set of video features representing properties of a video segment;

passing the set of video features into a convex hull estimation model to generate a set of bitrate-resolution pairs based on the set of video features, each bitrate-resolution pair in the set of bitrate-resolution pairs:

defining a bitrate; and

defining a resolution predicted to yield a quality score, approximating a maximum quality score, characterizing the video segment encoded at the bitrate and the resolution;

accessing a distribution of audience bandwidths representative of likely viewers of the video segment;

selecting a highest bitrate-resolution pair for the encoding ladder from the set of bitrate-resolution pairs, the highest bitrate-resolution pair defining a highest bitrate;

selecting a lowest bitrate-resolution pair for the encoding ladder from the set of bitrate-resolution pairs, the lowest bitrate-resolution pair defining a lowest bitrate;

selecting a subset of bitrate-resolution pairs from the set of bitrate-resolution pairs based on the distribution of audience bandwidths, the subset of bitrate-resolution pairs defining bitrates less than the highest bitrate and greater than the lowest bitrate; and

generating the encoding ladder, for the video segment, comprising the highest bitrate-resolution pair, the lowest bitrate-resolution pair, and the subset of bitrate-resolution pairs.

8. A method for generating an encoding ladder comprising:

extracting a set of video features representing properties of a video segment;

generating a feature vector for the video segment based on the set of video features;

generating a set of bitrate-resolution pairs via a convex hull estimation model and based on the feature vector for the video segment, each bitrate-resolution pair in the set of bitrate-resolution pairs:

defining a bitrate; and

defining a resolution predicted to yield a quality score, approximating a maximum quality score, characterizing the video segment encoded at the bitrate and the resolution;

accessing a distribution of audience bandwidths representative of likely viewers of the video segment;

selecting a highest bitrate-resolution pair for the encoding ladder from the set of bitrate-resolution pairs, the highest bitrate-resolution pair defining a highest bitrate;

selecting a lowest bitrate-resolution pair for the encoding ladder from the set of bitrate-resolution pairs, the lowest bitrate-resolution pair defining a lowest bitrate;

selecting a subset of bitrate-resolution pairs from the set of bitrate-resolution pairs based on the distribution of audience bandwidths, the subset of bitrate-resolution pairs defining bitrates less than the highest bitrate and greater than the lowest bitrate; and

generating the encoding ladder, for the video segment, comprising the highest bitrate-resolution pair, the lowest bitrate-resolution pair, and the subset of bitrate-resolution pairs.