Large Kernel Matters —— Improve Semantic Segmentation by Global Convolutional Network

Large Kernel Matters —— Improve Semantic Segmentation by Global Convolutional Network

8 Mar 2017 | Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, Jian Sun
The paper "Large Kernel Matters —— Improve Semantic Segmentation by Global Convolutional Network" by Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, and Jian Sun addresses the challenges of semantic segmentation, particularly the trade-off between classification and localization tasks. The authors propose a Global Convolutional Network (GCN) that combines the benefits of both classification and localization by using large kernels to enhance the receptive field and densely connect feature maps with per-pixel classifiers. They also introduce a Boundary Refinement block to improve object boundary localization. The GCN is designed to be fully convolutional, avoiding the use of fully-connected or global pooling layers that can discard localization information. The approach is evaluated on two public benchmarks: PASCAL VOC 2012 and Cityscapes, achieving state-of-the-art performance with 82.2% and 76.9% mean intersection-over-union (IoU) respectively. The paper includes extensive ablation studies and comparisons with other methods, demonstrating the effectiveness of large kernels and the proposed boundary refinement technique.The paper "Large Kernel Matters —— Improve Semantic Segmentation by Global Convolutional Network" by Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, and Jian Sun addresses the challenges of semantic segmentation, particularly the trade-off between classification and localization tasks. The authors propose a Global Convolutional Network (GCN) that combines the benefits of both classification and localization by using large kernels to enhance the receptive field and densely connect feature maps with per-pixel classifiers. They also introduce a Boundary Refinement block to improve object boundary localization. The GCN is designed to be fully convolutional, avoiding the use of fully-connected or global pooling layers that can discard localization information. The approach is evaluated on two public benchmarks: PASCAL VOC 2012 and Cityscapes, achieving state-of-the-art performance with 82.2% and 76.9% mean intersection-over-union (IoU) respectively. The paper includes extensive ablation studies and comparisons with other methods, demonstrating the effectiveness of large kernels and the proposed boundary refinement technique.
Reach us at info@study.space