This paper presents a novel approach to improve pixel-wise semantic segmentation by enhancing convolution operations. The authors introduce two key contributions: Dense Upsampling Convolution (DUC) and Hybrid Dilated Convolution (HDC). DUC is designed to generate pixel-level predictions by upscaling feature maps, capturing and decoding detailed information that is often lost in bilinear upsampling. HDC addresses the "gridding issue" in standard dilated convolution by using a range of dilation rates, effectively increasing the receptive field and improving accuracy for larger objects. The authors evaluate their methods on the Cityscapes dataset, achieving a state-of-the-art mIOU of 80.1% on the test set. They also report superior performance on the KITTI road estimation benchmark and the PASCAL VOC2012 segmentation task. The proposed techniques are integrated into a fully convolutional network (FCN) framework and can be easily implemented using existing deep learning tools.This paper presents a novel approach to improve pixel-wise semantic segmentation by enhancing convolution operations. The authors introduce two key contributions: Dense Upsampling Convolution (DUC) and Hybrid Dilated Convolution (HDC). DUC is designed to generate pixel-level predictions by upscaling feature maps, capturing and decoding detailed information that is often lost in bilinear upsampling. HDC addresses the "gridding issue" in standard dilated convolution by using a range of dilation rates, effectively increasing the receptive field and improving accuracy for larger objects. The authors evaluate their methods on the Cityscapes dataset, achieving a state-of-the-art mIOU of 80.1% on the test set. They also report superior performance on the KITTI road estimation benchmark and the PASCAL VOC2012 segmentation task. The proposed techniques are integrated into a fully convolutional network (FCN) framework and can be easily implemented using existing deep learning tools.