SubCost-Net
Rethinking the Stereo Matching in Light Field Disparity Estimation
1): Subpixel label is introduced for LF disparity estimation, which can significantly increase the cost compu-tation accuracy with the case of narrow baseline.
2): A novel subpixel cost computation block is designed by proposing different convolution operations with specific kernel size, dilation and padding region. With extensive ablation studies,
the effectiveness of the proposed block is validated.
full grid 9x9 angular views
Python
Nvidia RTX3090
