paritial derivative of L with respect to mean2D in backward.cu

[Xusj0w0]

I would like to ask about the computation of the partial derivative of the loss function L with respect to mean2D in backward.cu. There is an additional factor of 0.5W(ddelx_dx)in the computation.

atomicAdd(&dL_dmean2D[global_id].x, dL_dG * dG_ddelx * ddelx_dx);
atomicAdd(&dL_dmean2D[global_id].y, dL_dG * dG_ddely * ddely_dy);

I assume that if the quadratic form of the ellipse were computed in the NDC coordinate system, then multiplying by 0.5W would be reasonable. However, the computation above is performed in pixel coordinates, which confuses me.

[Wang-Chbo]

The calculation is the derivative of Loss w.r.t pixel coordinates