Grouped feature map quantisation
AssigneeCanon Kabushiki Kaisha
InventorsChristopher James Rosewarne, Jonathan GAN
Life of this patent
Estimating the term…
Abstract
A method of decoding feature maps from encoded data. A parameter related to quantisation is determined. According to the parameter, inverse quantisation is performed for samples decoded from the encoded data to decode the feature maps. For at least a part of the samples decoded from the encoded data, one parameter is shared by a group of samples corresponding to a plurality of feature maps.
Claims8
A method of decoding feature maps from a bitstream, the method comprising:
- decoding, from the bitstream, a frame including a first set of integer feature maps;
- decoding information specifying a maximum value of a first set of floating-point feature maps;
- deriving the first set of floating-point feature maps based on the maximum value and the first set of integer feature maps;
- and
- decoding a flag indicating whether the maximum value is updated,
- wherein when the flag indicates the maximum value is updated, an updated maximum value is used for deriving a second set of floating-point feature maps different from the first set of floating-point feature maps, and when the flag indicates the maximum value is not updated, the maximum value is used for deriving the second set of floating-point feature maps.
An apparatus for decoding feature maps from a bitstream, the apparatus comprising:
- one or more controllers including one or more processors and one or more memories, the one or more controllers configured to:
- decode, from the bitstream, a frame including a first set of integer feature maps;
- decode information specifying a maximum value of a first set of floating-point feature maps;
- derive the first set of floating-point feature maps based on the maximum value and the first set of integer feature maps;
- and
- decode a flag indicating whether the maximum value is updated,
- wherein when the flag indicates the maximum value is updated, an updated maximum value is used for deriving a second set of floating-point feature maps different from the first set of floating-point feature maps, and when the flag indicates the maximum value is not updated, the maximum value is used for deriving the second set of floating-point feature maps.
A non-transitory computer-readable storage medium which stores a program for executing a method decoding feature maps from data bitstream, the method comprising:
- decoding, from the bitstream, a frame including a first set of integer feature maps;
- decoding information specifying a maximum value of a first set of floating-point feature maps;
- deriving the first set of floating-point feature maps based on the maximum value and the first set of integer feature maps;
- and
- decoding a flag indicating whether the maximum value is updated,
- wherein when the flag indicates the maximum value is updated, an updated maximum value is used for deriving a second set of floating-point feature maps different from the first set of floating-point feature maps, and when the flag indicates the maximum value is not updated, the maximum value is used for deriving the second set of floating-point feature maps.
A method of encoding feature maps into encoded data, the method comprising:
- obtaining a set of floating-point feature maps;
- encoding a maximum value of the set of floating-point feature maps;
- deriving a first set of integer feature maps based on the set of floating-point feature maps and the maximum value;
- encoding a frame including the first set of integer feature maps into the bitstream;
- and
- encoding a flag indicating whether the maximum value is updated,
- wherein when the flag indicates the maximum value is updated, an updated maximum value is used for deriving a second set of integer feature maps different from the first set of integer feature maps, and when the flag indicates the maximum value is not updated, the maximum value is used for deriving the second set of integer feature maps.
An apparatus for encoding feature maps into a bitstream, the apparatus comprising:
- one or more controllers including one or more processors and one or more memories, the one or more controllers configured to:
- obtain a set of floating-point feature maps;
- encode a maximum value of the set of floating-point feature maps;
- derive a first set of integer feature maps based on the set of floating-point feature maps and the maximum value;
- encode a frame including the first set of integer feature maps into the bitstream;
- and
- encode a flag indicating whether the maximum value is updated,
- wherein when the flag indicates the maximum value is updated, an updated maximum value is used for deriving a second set of integer feature maps different from the first set of integer feature maps, and when the flag indicates the maximum value is not updated, the maximum value is used for deriving the second set of integer feature maps.
A non-transitory computer-readable storage medium which stores a program for executing a method encoding feature maps into encoded data, the method comprising:
- obtaining a plurality of floating-point feature maps;
- obtaining a set of floating-point feature maps;
- encoding a maximum value of the set of floating-point feature maps;
- deriving a first set of integer feature maps based on the set of floating-point feature maps and the maximum value;
- encoding a frame including the first set of integer feature maps into the bitstream;
- and
- encoding a flag indicating whether the maximum value is updated,
- wherein when the flag indicates the maximum value is updated, an updated maximum value is used for deriving a second set of integer feature maps different from the first set of integer feature maps, and when the flag indicates the maximum value is not updated, the maximum value is used for deriving the second set of integer feature maps.
Dependencies are inferred from the claim text.