Fast image scanning with deep max-pooling convolutional neural networks

By Alessandro Giusti, Dan Claudiu Ciresan, Jonathan Masci, Luca Maria Gambardella, J. Urgen Schmidhuber. 2013

360 citations recorded by the source corpus.

Deep Neural Networks now excel at image classification, detection and segmentation. When used to scan images by means of a sliding window, however, their high computational complexity can bring even the most powerful hardware to its knees. We show how dynamic programming can speedup the process by orders of magnitude, even when max-pooling layers are present.

Subfields: Computer Vision and Pattern Recognition, Electrical and Electronic Engineering, Efficiency & Compression

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 Electrical and Electronic Engineering

In Efficiency & Compression

Citations in the corpus

Fully convolutional networks for semantic segmentation (2014, methodology): “2 describes a trick, fast scanning [11], introduced for this purpose.”
OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks (2013, methodology): “To circumvent this problem, we take an approach similar to that introduced by Giusti et al. [9], and apply the last subsampling operation at every offset.”
DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs (2016, methodology): “We advocate instead the use of atrous convolution, originally developed for the efficient computation of the undecimated wavelet transform in the “algorithme a trous” scheme of [15] and used before in the DCNN context by [3], [6], [16].”
CrossTransformers: spatially-aware few-shot transfer (2020, methodology): “For experiments with CrossTransformers, we also increased the resolution of the convolutional feature map by setting the stride of final block of the ResNet to 1, and using dilated convolutions to preserve the feature computation [32, 39].”
Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs (2014, methodology): “To compute scores more densely at our target stride of 8 pixels, we develop a variation of the method previously employed by Giusti et al. (2013); Sermanet et al. (2013).”
Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation (2018, methodology): “In this section, we briefly introduce atrous convolution [30, 21, 64, 56, 8] and depthwise separable convolution [67, 71, 74, 12, 31].”
Rethinking Atrous Convolution for Semantic Image Segmentation (2017, methodology): “Instead, we advocate the use of ‘atrous convolution’, originally developed for the efficient computation of the undecimated wavelet transform in the “algorithme à trous” scheme of [30] and used before in the DCNN context by [21, 64, 56].”
Joint Training of a Convolutional Network and a Graphical Model for Human Pose Estimation (2014, background): “[13, 30] has addressed this problem by performing the convolution stages (with padding) on the full image input to efﬁciently create dense feature maps.”
Fully Connected Deep Structured Networks (2015, methodology): “for ﬁne-tuning. 4.1 Model Our model setup follows [3], i.e., we employ the 16 layer DeepNet model [31]. Just like [3] we ﬁrst convert the fully connected layers into convolutions as ﬁrst discussed in [12, 30]. This is useful since we are not interested in a single variable output prediction, but rat”
Deep learning in neural networks: An overview (2014, methodology): “they became the ﬁrst recurrent Deep Learners to win ofﬁcial inter national pattern recognition contests—see Sec. 5.17. 21 naive implementations by up to three orders of magnitude (Masci et al., 2013; Giusti et al., 2013); compare earlier efﬁcient methods for CNNs without MP (Vaillant et al., 1994). ”
Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Image Segmentation (2019, neutral): “Therefore, PSPNet [87] performs spatial pyramid pooling [21, 41, 24] at several grid scales (including image-level pooling [50]), while DeepLab [8, 9] applies several parallel atrous convolution [28, 20, 67, 57, 7] with different rates.”
Searching for Efficient Multi-Scale Architectures for Dense Image Prediction (2018, background): “Several works have focused on sampling rates in atrous convolution to encode multi-scale context [37, 29, 77, 62, 10, 94, 11].”
Monocular Object Instance Segmentation and Depth Ordering with CNNs (2015, methodology): “In contrast, we need a pixel-wise prediction, i.e., a categorical variable for each pixel. Therefore we ﬁrst convert the three fully-connected layers of the VGG network into convolutional units as in [8,33,25]. After this conversion, an increase of the size of the input patches to 306 306 translates”
MaskLab: Instance Segmentation by Refining Object Detection with Semantic and Direction Features (2017, methodology): “Atrous convolution: We apply the atrous convolution [34, 27, 64, 57], which has been successfully explored in semantic segmentation [13, 79, 12], object detection [19, 35] and instance segmentation [78, 44], to extract denser feature maps.”
DenseNet: Implementing Efficient ConvNet Descriptor Pyramids (2014, neutral): “compute CNN pyramids and perform sliding-window processing for image segmentation [18].”
Learning to Track for Spatio-Temporal Action Localization (2015, positive): “The slidingwindow procedure using CNN features can be performed efficiently [10, 36].”
Highly Efficient Forward and Backward Propagation of Convolutional Neural Networks for Pixelwise Classification (2014, methodology): “In contrast, fast scanning [7] still keeps large strides in its operations and is not ideal for GPU implementation.”
Real-Time Pedestrian Detection with Deep Network Cascades (2015, background): “Recently a novel range of methods have emerged, based on Deep Neural Networks, showing impressive accuracy gains [24]. c© 2015.”