Abstract: An image decoding method performed by a decoding apparatus according to the present document includes constituting a candidate list for deriving motion information of a sub-block unit for a current block, wherein the merge candidate list includes constructed candidates; deriving control point motion vectors (CPMVs) for control points (CPs) of the current block based on the merge candidate list; deriving prediction samples for the current block based on the CPMVs; and generating a reconstructed picture for the current block based on the derived prediction samples, wherein the constructed candidates are derived based on combination of at least two of a first neighboring block in a first group, a second neighboring block in a second group, a third neighboring block in a third group and a fourth neighboring block, and the first neighboring block, the second neighboring block and the third neighboring block and the fourth neighboring block constituting the combination have the same reference picture index.

Abstract: An image sensor module, an image processing device, and an image compression method receive pixel values of a target pixel group to be compressed in the image data and reference values of reference pixels to be used for compression of the target pixel group, generate a virtual reference map by applying an offset value to each of the reference values, compress the pixel values of the target pixel group based on the virtual reference map, and generate a bitstream based on a compression result and compression information based on the virtual reference map.

Abstract: A video decoding method performed by a decoding apparatus comprises deriving control points (CPs) for the current block; obtaining motion vectors for the CPs; deriving a motion vector of a sub-block or a sample unit in the current block on the basis of the obtained motion vectors; deriving a prediction sample for the current block on the basis of the derived motion vector; and generating a reconstruction sample on the basis of the prediction sample. The method enables effective performance of inter prediction through the motion vectors (transformation prediction), not only when an image in the current block is moved in a plane, but also when the image in the current block is rotated, zoomed in, zoomed out, or transformed into a parallelogram. Accordingly, the amount of data for the residual signal for the current block can be eliminated or reduced, and the overall coding efficiency can be improved.

Abstract: Techniques of compressing triangular mesh data involve generating a neighborhood table (i.e., a table) of fixed size that represents a neighborhood of a predicted vertex of a triangle within a triangular mesh for input into a machine-learning (ML) engine. For example, such a neighborhood table as input into a ML engine can output a prediction for a value (e.g., a position) of a vertex. The residual between the prediction and the actual value of the vertex is stored in an array. The data in the array representing the residuals may be compressed and transmitted over a network. Upon receipt by a computer, the array may be decompressed by the computer. Obtaining the actual value involves the receiving computer generating the same neighborhood table, inputting that neighborhood table into the same ML engine to produce the predicted value, and adding the predicted value to the residual from the decompressed file.

Abstract: An object recognition device includes a memory, a processor and a computer program stored in the memory and executed by the processor. When the computer program is executed by the processor, the processor implements the following steps: extracting object recognition feature vectors based on at least one frame image of the video on the first tracking target; matching the extracted object recognition feature vectors with object feature vectors of registered objects in a registration object library, the registration object library being established based on pre-learning of the object feature vectors in images of the registered objects; and taking one of the registered objects as a recognition result if the extracted object recognition feature vectors match with the object feature vectors of the one of the registered objects.

Abstract: A video tracking method comprises: determining a first tracking target in a video, and performing a video tracking process on the first tracking target based on the video; performing area overlap degree calculation on the first tracking target in a current image and the first tracking target in a past image during the video tracking process, the past image being an image obtained when the first tracking target was tracked last time in the video tracking process; extracting respective feature information of feature points of the first tracking target in the video during the video tracking process on the first tracking target; matching the extracted respective feature information of the feature points of the first tracking target in the current image with the respective feature information of corresponding feature points of the first tracking target in the past image if a result of the area overlap degree calculation is less than or equal to a preset threshold; and continuing the video tracking process on the f