This paper introduces two novel convolutional operations—Dense Upsampling Convolution (DUC) and Hybrid Dilated Convolution (HDC)—to improve semantic segmentation. DUC is designed to generate pixel-level predictions by learning an array of upscaling filters to upscale feature maps, which helps recover fine details that are often lost in bilinear upsampling. HDC addresses the "gridding" issue in dilated convolutions by using a range of dilation rates and concatenating them serially, which helps maintain a dense receptive field and improves the accuracy of large objects. The proposed methods are evaluated on the Cityscapes, KITTI road estimation, and PASCAL VOC2012 datasets, achieving state-of-the-art results. On the Cityscapes dataset, the DUC-HDC model achieves 80.1% mIOU on the test set. The methods are also effective in improving performance on other tasks, including road segmentation and object detection. The results show that DUC and HDC significantly enhance the accuracy of semantic segmentation by better capturing detailed information and maintaining a dense receptive field. The approach is implemented in a fully end-to-end trainable framework, and the source code is available at https://github.com/TuSimple/TuSimple-DUC.This paper introduces two novel convolutional operations—Dense Upsampling Convolution (DUC) and Hybrid Dilated Convolution (HDC)—to improve semantic segmentation. DUC is designed to generate pixel-level predictions by learning an array of upscaling filters to upscale feature maps, which helps recover fine details that are often lost in bilinear upsampling. HDC addresses the "gridding" issue in dilated convolutions by using a range of dilation rates and concatenating them serially, which helps maintain a dense receptive field and improves the accuracy of large objects. The proposed methods are evaluated on the Cityscapes, KITTI road estimation, and PASCAL VOC2012 datasets, achieving state-of-the-art results. On the Cityscapes dataset, the DUC-HDC model achieves 80.1% mIOU on the test set. The methods are also effective in improving performance on other tasks, including road segmentation and object detection. The results show that DUC and HDC significantly enhance the accuracy of semantic segmentation by better capturing detailed information and maintaining a dense receptive field. The approach is implemented in a fully end-to-end trainable framework, and the source code is available at https://github.com/TuSimple/TuSimple-DUC.