Abstract: The present disclosure provides pre-analysis based methods for adaptively compressing image data consumed by machine vision tasks. An exemplary method includes: receiving a video sequence; encoding one or more input pictures associated with the video sequence; and generating a bitstream, wherein the encoding includes: performing instance segmentation of an input picture, to generate one or more segment masks; combining the one or more segment masks to generate a merged mask; extracting, from the input picture, a region comprising the merged mask; and compressing image data representing the extracted region.

Abstract: An image data encoding method is provided. The image data encoding method includes receiving an input image and a quantization parameter (QP); performing pre-analysis of the input image based on the QP to obtain input image data for machine version; determining at least one of a temporal redundancy or a spatial redundancy of the input image data; determining, based on the at least one of temporal redundancy or spatial redundancy, whether to perform at least one of temporal down-sampling or spatial down-sampling; performing down-sampling of the input image data according to the determination, wherein the down-sampling includes at least one of temporal down-sampling or spatial down-sampling; and encoding the down-sampled image data based on the QP.

Abstract: A method for temporal resampling for multi-task machine vision is provided. The method includes receiving a bitstream of a video sequence after temporal resampling; and constructing a target frame from the bitstream using a frame construction model.

Abstract: A computer-implemented method for training spatial resampling modules includes: down-sampling, by a down-sampling module, an input image data to generate a down-sampled image data; up-sampling, by an up-sampling module, the down-sampled image data to generate a first up-sampled image data; analyzing, by a plurality of analysis models corresponding to a plurality of tasks, the first up-sampled image data; and training the down-sampling module based on a loss function associated with the plurality of analysis models according to the input image data and the first up-sampled image data.