Method and apparatus for generating video intermediate frame

1 . A method for generating a video intermediate frame, the method comprising:

performing a warp operation on a plurality of target video frames, based on a bidirectional optical flow between the plurality of target video frames, to obtain a plurality of pictures;

determining a similarity between sub-pictures corresponding to image sub-regions in the plurality of pictures;

predicting, based on the similarity, a network depth of a frame synthesis network matched with a corresponding image sub-region, the network depth increasing as the similarity decreases;

performing synthetic processing on corresponding sub-pictures of each of the image sub-regions using the frame synthesis network having the network depth matched with the corresponding image sub-region, to obtain a plurality of images; and

splicing the plurality of images to obtain intermediate frames of the plurality of target video frames, based on each of the image sub-regions.

2 . The method of claim 1 , wherein the performing of the warp operation comprises:

performing a forward-warping operation on the plurality of target video frames.

3 . The method of claim 1 , wherein a size of each of the image sub-regions is less than or equal to 64×64 dots-per-inch (dpi).

4 . The method of claim 1 , wherein the predicting of the network depth comprises:

predicting the network depth of the frame synthesis network based on the plurality of pictures and using a fully convolutional network (FCN).

5 . The method of claim 1 , wherein the frame synthesis network comprises a U-Net network.

6 . The method of claim 1 , wherein the performing of the synthetic processing comprises:

performing synthetic processing on the corresponding sub-pictures of a first image sub-region of the image sub-regions using the frame synthesis network with a first network depth matched to the first image sub-region; and

performing synthetic processing on the corresponding sub-pictures of a second image sub-region of the image sub-regions using the frame synthesis network with a second network depth matched to the second image sub-region,

wherein a first similarity of the corresponding sub-pictures of the first image sub-region is greater than a second similarity of the corresponding sub-pictures of the second image sub-region, and

wherein the first network depth is less than the second network depth.

7 . The method of claim 1 , wherein the performing of the warp operation comprises performing the warp operation using an optical flow estimation network,

wherein the predicting of the network depth comprises predicting the network depth of the frame synthesis network using a fully convolutional network (FCN), and

wherein the method further comprises integrating at least one of the optical flow estimation network, the FCN, and the frame synthesis network into an intermediate frame generation model.

8 . The method of claim 7 , further comprising:

training the intermediate frame generation model,

wherein the training of the intermediate frame generation model comprises:

providing sample video frames as input data to the intermediate frame generation model;

generating intermediate frames corresponding to the sample video frames;

calculating loss function values based on the intermediate frames corresponding to the sample video frames; and

adjusting, based on the loss function values, at least one network parameter of at least one of the optical flow estimation network, the FCN, and the frame synthesis network.

9 . A device for generating a video intermediate frame, comprising:

one or more processors;

a memory storing instructions that, when executed by the one or more processors, cause the device to:

perform a warp operation on a plurality of target video frames, based on a bidirectional optical flow between the plurality of target video frames, to obtain a plurality of pictures;

determine a similarity between sub-pictures corresponding to image sub-regions in the plurality of pictures;

predict, based on the similarity, a network depth of a frame synthesis network matched with a corresponding image sub-region, the network depth increasing as the similarity decreases;

perform synthetic processing on corresponding sub-pictures of each of the image sub-regions using the frame synthesis network having the network depth matched with the corresponding image sub-region, to obtain a plurality of images; and

splice the plurality of images to obtain intermediate frames of the plurality of target video frames, based on each of the image sub-regions.

10 . The device of claim 9 , wherein the instructions, when executed by the one or more processors, further cause the device to:

perform a forward-warping operation on the plurality of target video frames.

11 . The device of claim 9 , wherein a size of each of the image sub-regions is less than or equal to 64×64 dots-per-inch (dpi).

12 . The device of claim 9 , wherein the instructions, when executed by the one or more processors, further cause the device to:

predict the network depth of the frame synthesis network based on the plurality of pictures and using a fully convolutional network (FCN).

13 . The device of claim 9 , wherein the frame synthesis network comprises a U-Net network.

14 . The device of claim 9 , wherein the instructions, when executed by the one or more processors, further cause the device to:

perform synthetic processing on the corresponding sub-pictures of a first image sub-region of the image sub-regions using the frame synthesis network with a first network depth matched to the first image sub-region; and

perform synthetic processing on the corresponding sub-pictures of a second image sub-region of the image sub-regions using the frame synthesis network with a second network depth matched to the second image sub-region,

wherein a first similarity of the corresponding sub-pictures of the first image sub-region is greater than a second similarity of the corresponding sub-pictures of the second image sub-region, and

wherein the first network depth is less than the second network depth.

15 . The device of claim 9 , wherein the instructions, when executed by the one or more processors, further cause the device to:

perform the warp operation using an optical flow estimation network;

predict the network depth of the frame synthesis network using a fully convolutional network (FCN), and

integrate at least one of the optical flow estimation network, the FCN, and the frame synthesis network into an intermediate frame generation model.