Dilated Residual Networks

By Fisher Yu, Vladlen Koltun, Thomas A. Funkhouser. 2017

1,817 citations recorded by the source corpus.

Convolutional networks for image classification progressively reduce resolution until the image is represented by tiny feature maps in which the spatial structure of the scene is no longer discernible. Such loss of spatial acuity can limit image classification accuracy and complicate the transfer of the model to downstream applications that require detailed scene understanding. These problems can be alleviated by dilation, which increases the resolution of output feature maps without reducing the receptive field of individual neurons. We show that dilated residual networks (DRNs) outperform their non-dilated counterparts in image classification without increasing the models depth or complexity. We then study gridding artifacts introduced by dilation, develop an approach to removing these artifacts (degridding), and show that this further increases the performance of DRNs. In addition, we show that the accuracy advantage of DRNs is further magnified in downstream applications such as object localization and semantic segmentation.

Subfields: Artificial Intelligence, Biophysics, Computer Vision and Pattern Recognition, Model Architectures

How researchers see this paper

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

In Artificial Intelligence

In Biophysics

In Computer Vision and Pattern Recognition

In Model Architectures

Citations in the corpus

Making Convolutional Networks Shift-Invariant Again (2019, methodology): “Deep networks can also be evaluated densely, by removing striding and making appropriate changes to subsequent layers by using ´a trous /dilated convolutions (Chen et al., 2015; 2018; Yu & Koltun, 2016; Yu et al., 2017).”
UPSNet: A Unified Panoptic Segmentation Network (2019, background): “As context is crucial for semantic segmentation, one notable improvement to most convolutional models emerged from dilated convolutions [37, 38] which allows for a larger receptive field without the need of more free parameters.”
Generative Modeling by Estimating Gradients of the Data Distribution (2019, methodology): “In the experiments, our model sθ(x, σ) combines the architecture design of U-Net [46] with dilated/atrous convolution [64, 65, 8]—both of which have been proved very successful in semantic segmentation.”
Semantic Image Synthesis With Spatially-Adaptive Normalization (2019, methodology): “We use the state-of-the-art segmentation networks for each dataset: DeepLabV2 [5, 40] for COCO-Stuff, UperNet101 [51] for ADE20K, and DRN-D-105 [53] for Cityscapes.”
Deformable Convolutional Networks (2017, methodology): “It has been widely used for semantic segmentation [41, 5, 54] (also called dilated convolution in [54]), object detection [7], and image classiﬁcation [55].”
Why do deep convolutional networks generalize so poorly to small image transformations? (2018, background): “This does not seem feasible for larger size images (although see (Chen et al., 2014, 2017; Yu and Koltun, 2015; Yu et al., 2017) for ways to make this more efficient).”
Dilated Recurrent Neural Networks (2017, background): “All the tasks are sequence level classiﬁcation tasks, and therefore the “gridding” problem [29] is irrelevant.”
Evaluating Protein Transfer Learning with TAPE (2019, methodology): “Architectures and Training: We implement three architectures: an LSTM [48], a Transformer [49], and a dilated residual network (ResNet) [50].”
Continual Semantic Segmentation via Repulsion-Attraction of Sparse and Disentangled Latent Representations (2021, background): “Current approaches are based on fully convolutional models [38, 7, 6, 75, 72] and exploit various techniques to cope with multi-scale and spatial dependency.”
Knowledge Distillation for Incremental Learning in Semantic Segmentation (2019, methodology): “Current state-of-the-art approaches are mostly based on the Fully Convolutional Network (FCN) model [15] and some of the most succesful are DRN [16], PSPNet [17] and DeepLab [18].”
Deep High-Resolution Representation Learning for Visual Recognition (2019, methodology): “The fully convolutional network is extended, by re-placing a few (typically two) strided convolutions and the associated convolutions with dilated convolutions, to the dilation version, leading to medium-resolution representations [18], [19], [86], [168], [181].”
CyCADA: Cycle-Consistent Adversarial Domain Adaptation (2017, background): “, 2015) base network and (B) DRN-26 (Yu et al., 2017).”
Context Encoding for Semantic Segmentation (2018, methodology): “We follow the prior work using dilated network strategy on pre-trained network [6, 55, 59] at stage 3 and 4 2 , as shown in Figure 3.”
Deep Layer Aggregation (2017, methodology): “Layerwise accuracy comparisons [11, 48, 35], transferability analysis [44], and representation visualization [48, 46] show that deeper layers extract more semantic and more global features, but these signs do not prove that the last layer is the ultimate representation for any task.”
BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning (2018, methodology): “Then, we train DRN-D-22 model [35] on the 70,000 training images.”
On the Computational Inefficiency of Large Batch Sizes for Stochastic Gradient Descent (2018, neutral): “We also test how these results vary across other modern DNN architectures, namely ResNets [10], LSTMs [11, 8], AlexNet [20], VGG [27], Dilated Residual Networks [34], and MobileNetV2 [25].”
Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells (2018, background): “lated convolutions for keeping the receptive field size intact [6, 44, 46].”
From Lifestyle Vlogs to Everyday Interactions (2017, methodology): “We modify a standard dilated Resnet-54 [48] (details in supplemental) as follows: we introduce a 2-layer network that maps δ to feature maps; the base feature map and δ feature map are then concatenated and fused by 3 convolutional layers to predict hand segmentation.”
DetNet: A Backbone network for Object Detection (2018, background): “Another different research is Dilated Residual Network [26] which extracts features with less strides.”