Large Kernel Matters — Improve Semantic Segmentation by Global Convolutional Network

By Chao Peng, Xiangyu Zhang, Gang Yu, Guozhong Luo. 2017

1,617 citations recorded by the source corpus.

One of recent trends [31, 32, 14] in network architecture design is stacking small filters (e.g., 1x1 or 3x3) in the entire network because the stacked small filters is more efficient than a large kernel, given the same computational complexity. However, in the field of semantic segmentation, where we need to perform dense per-pixel prediction, we find that the large kernel (and effective receptive field) plays an important role when we have to perform the classification and localization tasks simultaneously. Following our design principle, we propose a Global Convolutional Network to address both the classification and localization issues for the semantic segmentation. We also suggest a residual-based boundary refinement to further refine the object boundaries. Our approach achieves state-of-art performance on two public benchmarks and significantly outperforms previous results, 82.2% (vs 80.2%) on PASCAL VOC 2012 dataset and 76.9% (vs 71.8%) on Cityscapes dataset.

Subfields: Computer Vision and Pattern Recognition, Media Technology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Media Technology

Citations in the corpus

Axial-DeepLab: Stand-Alone Axial-Attention for Panoptic Segmentation (2020, methodology): “Some employ atrous convolutions [34,77,67,13], larger kernel [70], or image pyramids [98,85], either designed by hand or searched by algorithms [103,12,60].”
CCNet: Criss-Cross Attention for Semantic Segmentation (2018, methodology): “Most of methods [5, 22, 40, 27, 30, 41, 35, 19, 42, 36] adopt the same backbone as ours and the others [32, 34] utilize stronger backbones.”
MegDet: A Large Mini-Batch Object Detector (2017, methodology): “Based on our MegDet, we integrate the techniques including OHEM [35], atrous convolution [40, 2], stronger base models [38, 18], large kernel [28], segmentation supervision [27, 34], diverse network structure [12, 32, 36], contextual modules [22, 9], ROIAlign [14] and multi-scale training and testin”
FocalClick: Towards Practical Interactive Image Segmentation (2022, background): “The Segmentor could be any segmentation network [4, 28, 32, 41, 47] for customized scenarios.”
Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation (2018, methodology): “Encoder-decoder: The encoder-decoder networks have been successfully applied to many computer vision tasks, including human pose estimation [53], object detection [45, 66, 19], and semantic segmentation [49, 54, 61, 3, 43, 59, 57, 33, 76, 20].”
Rethinking Atrous Convolution for Semantic Image Segmentation (2017, methodology): “Deconvolutional layers (or transposed convolution) [79, 51, 54, 3, 61, 58] have been employed to recover the spatial resolution.”
Hybrid Task Cascade for Instance Segmentation (2019, methodology): “GCN.”
Path Aggregation Network for Instance Segmentation (2018, background): “[47] where global convolutions were utilized.”
ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design (2018, background): “This motivates a practical question: how to increase the size of receptive field? This is critical for object detection in high-resolution images [20].”
DetNAS: Backbone Search for Object Detection (2019, methodology): “ow-resolution, semantically strong features with highresolution, semantically weak features. The feature pyramid expand the effect of each individual backbone stage. In addition to DetNAS, other work [18] has also made detailed analyses on kernel size in the task of performing the classiﬁcation and ”
Representation Compensation Networks for Continual Semantic Segmentation (2022, methodology): “Currently methods pay more attention to multi-scale feature aggregation [3, 34, 52,53,59,65,68,81].”
Searching for Efficient Multi-Scale Architectures for Dense Image Prediction (2018, methodology): “…objects are captured by the downsampled image [26, 66, 23, 50, 13, 11], (2) designing models that contain encoder-decoder structures [2, 72, 49, 28, 64, 93, 96], or (3) designing models that employ a multi-scale context module, e.g ., DenseCRF module [42, 4, 10, 98, 50, 76], global context [56,…”
Dual Attention Network for Scene Segmentation (2018, background): “However, many works [15, 29] suggest that local features generated by traditional FCNs could lead to misclassification of objects and stuff.”
Deep High-Resolution Representation Learning for Visual Recognition (2019, background): “…and heavy up-sample processes [141], recombinator networks [55]; improving skip connections with more or complicated convolutional units [64], [111], [180], as well as sending information from low-resolution skip connections to high-resolution skip connections [189] or exchanging information…”
SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers (2021, background): “For semantic segmentation, maintaining large receptive field to include context information has been a central issue [5, 19, 20].”
ExFuse: Enhancing Feature Fusion for Semantic Segmentation (2018, neutral): “At the beginning of Sec 3 we demonstrate that feature fusion in our baseline architecture (GCN [8]) is ineffective.”
Strip Pooling: Rethinking Spatial Pooling for Scene Parsing (2020, methodology): “For continuous feature space learning, prior work use multi-scale feature aggregation [37, 5, 33, 18, 42, 31, 32, 2, 44, 4, 48, 17] to fuse the contextual information by probing the incoming features with filters or pooling operations at multiple rates and multiple fields-of-view.”