Transformer based video coding
An example method includes: receiving a target frame and one or more reference frames; extracting a first feature map from a first predicted target frame predicted from a first reference frame, and a second feature map from a second predicted frame predicted from a second target frame, wherein the first predicted target frame is a backward predicted target frame and the second predicted target frame is a forward predicted target frame; generating a refined residual feature based at least on the first feature map, the second feature map, and a third feature map extracted from a feature decoder net module or circuit; generating a frame residual based at least on the refined residual feature; and generating an output reconstructed frame based at least on the frame residual and an average frame, wherein the average frame represents an average of the first predicted target frame and the second predicted target frame.
1 . An apparatus comprising at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to perform:
receiving a target frame and one or more reference frames;
extracting three or more feature tensors at one or more different scales from each reference frame and the target frame;
providing a first set of feature tensors for first set of scales of the one or more different scales as an input to an attention block or circuit;
generating a second set of smaller feature tensors, as compared to the three or more feature tensors, for a second set of scales of the one or more different scales by splitting or pixel interleaving a second set of feature tensors, wherein the second set of smaller feature tensors are randomly shuffled, and wherein every 2 smaller feature tensors are combined to generate a first downscaled feature map; and
generating a third set of smaller feature tensors, as compared to the three or more feature tensors, for a third set of scales of the one or more different scales by splitting or pixel interleaving a third set of feature tensors, wherein the third set of smaller feature tensors are randomly shuffled, and wherein every 4 smaller feature tensors are combined to generate a second downscaled feature map.
2 . The apparatus of claim 1 , wherein the apparatus is further caused to generate an i th set of smaller feature tensors, as compared to the three or more feature tensors, for an i th set of scales of the one or more different scales by splitting or pixel interleaving an i th set of feature tensors, wherein the i th set of smaller feature tensors are randomly shuffled, and wherein every i smaller feature tensors are combined to generate an i th downscaled feature map, and wherein i is a natural number greater than 0.
3 . The apparatus of claim 1 , wherein the apparatus further comprises:
an encoder feature pyramid net block or circuit, wherein the three or more feature tensors are extracted by the encoder feature pyramid net block or circuit; and
a feature encoder net block or circuit, and wherein the apparatus is further caused to perform:
providing the extracted three or more feature tensors as an input to the feature encoder net block or circuit.
4 . The apparatus of claim 3 , wherein the feature encoder net block or circuit comprises one or more transformer encoders and/or one or more transformer decoders.
5 . The apparatus of claim 4 , wherein feature tensors at same scale and extracted from different reference frames and the target frame are provided as an input to the same transformer encoder, and wherein the transformer encoder is caused to perform: combining information between global features and local features.
6 . The apparatus of claim 4 , wherein each feature tensor has same dimension after splitting, and wherein the each split feature tensor works as memory signal for the one or more transformer decoders.
7 . The apparatus of claim 4 , wherein the first set of feature tensors of the target frame are provided as target for the one or more transformer decoders, and wherein the one or more transformer decoders are caused to perform: providing a latent tensor comprising the same dimension as the first set of feature tensors as output, and wherein the latent tensor is lossless encoded.
8 . The apparatus of claim 1 , wherein the three or more feature tensors comprise three or more three dimensional (3D) feature tensors, and wherein the apparatus further comprises a channel excitation multi-head attention block or circuit, and wherein the channel excitation multi-head attention block or circuit is caused to perform: computing an attention between the 3D feature tensors by performing:
channel excitation operation that calculates an average value of patches in a 3D feature tensor;
assigning different importance to each channel of the 3D feature tensor;
aggregating spatial feature information of the 3D feature tensor to a single channel;
passing the aggregated 3D feature tensor through a plurality of convolution layers to update dimension of the aggregated 3D feature tensor to one dimensional (1D) feature tensor; and
computing an attention map between the 1D feature tensor and 3D tensor feature maps by using a scaled dot product attention.
9 . The apparatus of claim 1 , wherein the apparatus further comprises a transformer comprising a transformer encoder and a transformer decoder, wherein the transformer receives following as input: features of a first scale of the one or more different scales; and output from a previous transformer layer, when the previous transformer layer is present; and wherein a feature map of the target frame is provided as an input to the transformer decoder, wherein the apparatus is further caused to perform:
sequentially processing the one or more different scales;
providing a latent tensor as an output; and
encoding the latent tensor by using a lossless encoder.
10 . The apparatus of claim 1 , wherein the one or more different scales comprise spatial resolution.
11 . An apparatus comprising at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to perform:
extracting a first feature map from a first predicted target frame predicted from a first reference frame, and a second feature map from a second predicted target frame predicted from a second reference frame, wherein the first predicted target frame is a backward predicted target frame and the second predicted target frame is a forward predicted target frame;
generating a refined residual feature based at least on the first feature map, the second feature map, and a third feature map extracted from a feature decoder net module or circuit;
generating a frame residual based at least on the refined residual feature; and
generating an output reconstructed frame based at least on the frame residual and an average frame, wherein the average frame represents an average of the first predicted target frame and the second predicted target frame.
12 . The apparatus of claim 11 , wherein the apparatus further comprises a whole transformer module or circuit comprising a transformer encoder and transformer decoder, and wherein the apparatus is further caused to perform: providing the first feature map and the second feature map to the whole transformer module or circuit to generate the refined residual feature, and wherein the third feature map represents a target to the transformer decoder.
13 . A method comprising:
receiving a target frame and one or more reference frames;
extracting three or more feature tensors at one or more different scales from each reference frame and the target frame;
providing a first set of feature tensors for first set of scales of the one or more different scales as an input to an attention block or circuit;
generating a second set of smaller feature tensors, as compared to the three or more feature tensors, for a second set of scales of the one or more different scales by splitting or pixel interleaving a second set of feature tensors, wherein the second set of smaller feature tensors are randomly shuffled, and wherein every 2 smaller feature tensors are combined to generate a first downscaled feature map; and
generating a third set of smaller feature tensors, as compared to the three or more feature tensors, for a third set of scales of the one or more different scales by splitting or pixel interleaving a third set of feature tensors, wherein the third set of smaller feature tensors are randomly shuffled, and wherein every 4 smaller feature tensors are combined to generate a second downscaled feature map.
14 . The method of claim 13 , wherein the three or more feature tensors are extracted by a feature encoder net block or circuit, and wherein the feature encoder net block or circuit comprises one or more transformer encoders and/or one or more transformer decoders.
15 . The method of claim 14 , wherein each feature tensor has same dimension after splitting, and wherein the each split feature tensor works as memory signal for the one or more transformer decoders.
16 . The method of claim 14 further comprising:
providing the first set of feature tensors of the target frame as target for the one or more transformer decoders; and
providing a latent tensor comprising the same dimension as the first set of feature tensors as output by the one or more transformer decoders, wherein the latent tensor is lossless encoded.
17 . The method of claim 13 , wherein the three or more feature tensors comprise three or more three dimensional (3D) feature tensors, and wherein the method further comprises computing an attention between the 3D feature tensors, wherein computing the attention between 3D feature tensors comprises:
performing a channel excitation operation that calculates an average value of patches in a 3D feature tensor;
assigning different importance to each channel of the 3D feature tensor;
aggregating spatial feature information of the 3D feature tensor to a single channel;
passing the aggregated 3D feature tensor through a plurality of convolution layers to update dimension of the aggregated 3D feature tensor to one dimensional (1D) feature tensor; and
computing an attention map between the 1D feature tensor and 3D tensor feature maps by using a scaled dot product attention.
18 . The method of claim 13 further comprising:
receiving following as input: features of a first scale of the one or more different scales; and output from a previous transformer layer, when the previous transformer layer is present, wherein a feature map of the target frame is provided as an input to a transformer decoder;
sequentially processing the one or more different scales;
providing a latent tensor as an output; and
encoding the latent tensor by using a lossless encoder.
19 . A method comprising:
extracting a first feature map from a first predicted target frame predicted from a first reference frame, and a second feature map from a second predicted target frame predicted from a second reference frame, wherein the first predicted target frame is a backward predicted target frame and the second predicted target frame is a forward predicted target frame;
generating a refined residual feature based at least on the first feature map, the second feature map, and a third feature map extracted from a feature decoder net module or circuit;
generating a frame residual based at least on the refined residual feature; and
generating an output reconstructed frame based at least on the frame residual and an average frame, wherein the average frame represents an average of the first predicted target frame and the second predicted target frame.
20 . The method of claim 19 further comprising: providing the first feature map and the second feature map to a whole transformer module or circuit, comprising a transformer encoder and transformer decoder, to generate the refined residual feature, wherein the third feature map represents a target to the transformer decoder.