keyword
spatial context model
A spatial context model is a mechanism in learned data and image compression that estimates the probability distribution of a target symbol or latent feature based on the values of its neighboring elements across spatial dimensions. By analyzing adjacent positions along the height and width of a feature representation, the model captures local statistical correlations and spatial redundancies that persist after initial feature transforms. These models typically employ autoregressive structures, masked convolutions, or multi-pass schemes such as checkerboard patterns to ensure that conditional predictions rely strictly on previously reconstructed spatial neighbors. By supplying these refined conditional probability distributions to an entropy coder, a spatial context model minimizes the required bit rate and significantly improves overall rate-distortion performance.
1 item
