Pixelwise Dynamic Convolution Neural Network for LiDAR Depth Data Interpolation

Deformable element-wise dynamic convolution

“…To enhance the flexibility of convolution during the test phase, a content-aware kernel, or a dynamic kernel, proposed. 6,7,[15][16][17]23,24 A kernel-generating method is one of the straightforward dynamic kernels. [15][16][17]23 Ha et al 15 introduced a compact network that generates kernel values for the target layer using input data, the target layer's location, and the kernel values of the previous layer.…”

“…There is another major approach, known as the kernel-combination method, 6,7,24 that has gained popularity in recent years. Yang et al 24 proposed a kernel-combination method called CondConv for convolution.…”

“…Chen's method represents the attention of each base kernel as a scalar value and generates a dynamic kernel at the data-level, resulting in what is referred to as a scene-wise dynamic convolution. By contrast, Kim et al 7 proposed a pixel-wise dynamic convolution approach that generates a dynamic kernel at the spatial-level and uses different kernel values in the spatial dimension to process input data. These kernel-combination methods were proven to be effective in many tasks and were much easier to optimize than kernelgenerating methods.…”

“…the proposed convolution method is able to flexibly process data more than the existing deformable convolution method 5 and dynamic convolution methods. 6,7 4 Experiments…”

“…As a solution, kernelcombination methods have garnered attention, 6,7 wherein multiple kernels are combined to generate new kernel values at either the data level 6 or spatial level. 7 These kernel-combination methods reduced optimization difficulty compared with kernel-generating methods and were effective across various tasks.…”