pacnet

Why we included this project

Researchers working on segmentation, depth estimation, and other dense prediction tasks will find this the most direct way to test the Pixel-Adaptive Convolution (PAC) layers introduced at CVPR 2019. It is the authors' reference implementation, and it ships five PyTorch modules: the standard convolution, a transposed variant for upsampling, a pooling layer, and two CRF mean-field inference implementations, all accepting the familiar nn.Conv2d arguments. Because PAC layers learn filter weights from pixel content instead of fixing them in advance, they give networks spatial adaptivity without much extra machinery. The code is compact enough to read in one sitting, and the included unit tests help you confirm the layers behave before wiring them into your own model. One honest caveat: the project is under a CC BY-NC-SA license and pins fairly old PyTorch versions.