Method and apparatus for action recognition

The method of 4, wherein the performing of the two cross-resolution convolution operations comprises:

• applying a temporal upsampling operation to the second feature tensor;
• and
• convolving the temporally upsampled second feature tensor with a 3×3 kernel;
• wherein the each TCR block combines the convolved second feature tensor with the first feature tensor using element-wise summation.

An apparatus for performing video action recognition, the apparatus comprising:

• a machine learning (ML) model for identifying actions in videos, the ML model comprising a plurality of temporal cross-resolution (TCR) blocks;
• a gallery application for viewing the videos on the apparatus;
• and
• at least one processor coupled to memory and configured to:
• receive a video comprising a plurality of frames;
• generate a multi-dimensional feature tensor representing the received video, the tensor having a plurality of channels;
• and
• perform action recognition using the ML model by:
• splitting the generated feature tensor into a first feature tensor having a first set of channels, and a second feature tensor having a second set of channels;
• applying at least one temporal pooling layer to the second feature tensor to temporally downsample the second feature tensor;
• processing, in parallel, using each TCR block among the plurality of TCR blocks:
• the first feature tensor, and
• the temporally downsampled second feature tensor;
• and
• outputting a prediction of a first action within the received video.