Motion compensation along different directions

1. A method for video processing, comprising:

determining, for a conversion between a current video block of a video and a bitstream of the current video block, an optical flow associated with the current video block in an optical flow-based motion refinement process or prediction process, wherein the optical flow is derived along directions that are different from a horizontal direction and a vertical direction; and

performing the conversion based on the optical flow.

2. The method of claim 1 , wherein spatial gradients associated with the current video block are derived along the same directions used for deriving the optical flow.

3. The method of claim 2 , wherein prediction refinements associated with the current video block are generated using the optical flow and the spatial gradients derived in the directions.

4. The method of claim 2 , wherein the optical flow or/and the spatial gradients are derived along a diagonal direction and an anti-diagonal direction,

wherein the diagonal direction refers to a horizontal direction rotated by M-degree anticlockwise, and

wherein the anti-diagonal direction refers to a vertical direction rotated by N-degree anticlockwise, M and N being integers.

5. The method of claim 4 , wherein M and/or N is equal to 45.

6. The method of claim 2 , wherein the optical flow or/and the spatial gradients are derived for one direction pair, and

wherein the one direction pair includes two directions which include horizontal and vertical directions or diagonal and anti-diagonal directions.

7. The method of claim 1 , wherein one or multiple prediction refinements associated with the current video block are generated in the optical flow-based motion refinement process or prediction process, and

wherein a final prediction refinement associated with the current video block is generated by combining the multiple prediction refinements.

8. The method of claim 7 , wherein the multiple prediction refinements are derived in multiple directions or multiple direction pairs,

wherein a first prediction refinement of the multiple prediction refinements is derived in a horizontal-vertical direction pair including horizontal and vertical directions, and a second prediction refinement of the multiple prediction refinements is derived in a diagonal-anti-diagonal direction pair including diagonal and anti-diagonal directions,

wherein the first prediction refinement for a reference picture list X is defined as:

ofX h ( x,y )×gradX h ( x,y )+of X v ( x,y )×gradX y ( x,y ),

where X=0 or 1, ofX h (x, y) and ofX v (x, y) denote a horizontal optical flow and a vertical optical flow for the reference picture list X respectively, gradX h (x, y) and gradX v (x, y) denote a horizontal gradient and a vertical gradient of PX(x, y), and PX(x, y) denotes a prediction sample of a sample (x, y) in the reference picture list X, and

wherein the second prediction refinement for the reference picture list X (X=0, 1) is defined as:

ofX d ( x,y )×gradX d ( x,y )+ ofX ad ( x,y )×gradX ad ( x,y ),

where X=0 or 1, ofX d (x, y) and of X ad (x, y) denote a diagonal optical flow and an anti-diagonal optical flow in the reference picture list X respectively, and gradX d (x, y) and gradX ad (x, y) denote a diagonal gradient and an anti-diagonal gradient of PX(x, y).

9. The method of claim 7 , wherein the multiple prediction refinements are weighted averaged to generate the final prediction refinement,

wherein weights of the multiple prediction refinements depend on gradient information of a prediction block associated with the current video block, and

wherein spatial gradients are calculated for multiple direction pairs and smaller weights are assigned to a direction pair with smaller spatial gradients, or spatial gradients are calculated for the multiple direction pairs and smaller weights are assigned to a direction pair with larger spatial gradients.

10. The method of claim 9 , wherein the weight for a first sample in a first prediction refinement block associated with the current video block is different from a second sample in the first prediction refinement block,

wherein default weights are assigned to the multiple prediction refinements,

wherein the final prediction refinement is generated for each reference picture list X, or

wherein the weights used for the multiple prediction refinements depend on reliability of multiple optical flows associated with the current video block.

11. The method of claim 1 , wherein an interpolation for a motion vector associated with the current video block is performed to generate an interpolation result in the optical flow-based motion refinement process or prediction process, and wherein the interpolation is performed along the directions that are different from the horizontal direction and/or the vertical direction.

12. The method of claim 11 , wherein the interpolation is performed along two directions orthogonal to each other, and the two directions are different from the horizontal direction and the vertical direction, or

wherein the interpolation is performed along a diagonal direction or/and an anti-diagonal direction, where the diagonal direction refers to a horizontal direction rotated by M-degree anticlockwise, and the anti-diagonal direction refers to a vertical direction rotated by N-degree anticlockwise, M and N being integers.

13. The method of claim 1 , wherein an interpolation for a motion vector associated with the current video block is performed to generate one or multiple interpolation results in the optical flow-based motion refinement process or prediction process, and

wherein a final interpolation result associated with the current video block is generated by combining the multiple interpolation results.

14. The method of claim 13 , wherein the multiple interpolation results are derived in multiple directions or direction pairs,

wherein a first interpolation result of the multiple interpolation results is generated in a horizontal-vertical direction pair including horizontal and vertical directions, and a second interpolation result of the multiple interpolation results is derived in a diagonal-anti-diagonal direction pair including diagonal and anti-diagonal directions, and

wherein the multiple interpolation results are weighted averaged to generate the final interpolation result.

15. The method of claim 14 , wherein the weights depend on gradient information of a reference block associated with the current video block,

wherein the weight for a first sample in a first interpolated block is different from a second sample in the first interpolated block,

wherein the weights are derived for each sample, or

wherein the weights are derived for each block or sub-block.

16. The method of claim 1 , wherein the optical flow-based motion refinement process or prediction refinement process is prediction refinement with optical flow (PROF) or bi-directional optical flow (BDOF).

17. The method of claim 1 , wherein the conversion includes encoding the current video block into the bitstream.

18. The method of claim 1 , wherein the conversion includes decoding the current video block from the bitstream.

19. An apparatus for processing video data comprising a processor and a non-transitory memory with instructions thereon, wherein the instructions upon execution by the processor, cause the processor to:

determine, for a conversion between a current video block of a video and a bitstream of the current video block, an optical flow associated with the current video block in an optical flow-based motion refinement process or prediction process, wherein the optical flow is derived along directions that are different from a horizontal direction and a vertical direction; and

perform the conversion based on the optical flow.

20. A non-transitory computer-readable recording medium storing a bitstream of a video which is generated by a method performed by a video processing apparatus, wherein the method comprises:

determining an optical flow associated with a current video block of the video in an optical flow-based motion refinement process or prediction process, wherein the optical flow is derived along directions that are different from a horizontal direction and a vertical direction; and

generating the bitstream based on the optical flow.