Densely Connected Convolutional Networks

Recent work has shown that convolutional networks can be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convolutional Network (DenseNet), which connects each layer to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections - one between each layer and its subsequent layer - our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less computation to achieve high performance. Code and pre-trained models are available at https://github.com/liuzhuang13/DenseNet .

PDF Abstract CVPR 2017 PDF CVPR 2017 Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Image Classification CIFAR-10 DenseNet (DenseNet-BC-190) Percentage correct 96.54 # 97
Top-1 Accuracy 96.54 # 17
Image Classification CIFAR-100 DenseNet-BC Percentage correct 82.82 # 96
Image Classification CIFAR-100 DenseNet Percentage correct 82.62 # 99
Image Classification GasHisSDB DenseNet-169 Accuracy 96.90 # 8
Precision 99.91 # 7
F1-Score 98.38 # 8
Image Classification ImageNet DenseNet-121 Top 1 Accuracy 74.98% # 890
Image Classification ImageNet DenseNet-169 Top 1 Accuracy 76.2% # 851
Image Classification ImageNet DenseNet-264 Top 1 Accuracy 77.85% # 794
Image Classification ImageNet DenseNet-201 Top 1 Accuracy 77.42% # 808
Classification InDL DenseNet201 Average Recall 90.99% # 3
Medical Image Classification NCT-CRC-HE-100K DenseNet-169 Accuracy (%) 94.41 # 5
F1-Score 96.90 # 5
Precision 99.87 # 7
Specificity 99.30 # 5
Image Classification SVHN DenseNet Percentage error 1.59 # 16
Pedestrian Attribute Recognition UAV-Human DenseNet Gender 75.0 # 1
Hat 67.2 # 1
UCC 49.8 # 1
UCS 73.0 # 1
LCC 54.6 # 1
LCS 68.9 # 2
Backpack 63.9 # 1
Classification XImageNet-12 DenseNet121 Robustness Score 0.9062 # 1

Results from Other Papers


Task Dataset Model Metric Name Metric Value Rank Uses Extra
Training Data
Source Paper Compare
Breast Tumour Classification PCam DenseNet-121 (e) AUC 0.921 # 12
Crowd Counting UCF-QNRF Densenet201 MAE 163 # 13

Methods