DenseNet: Implementing Efficient ConvNet Descriptor Pyramids

By Forrest N. Iandola, Matthew W. Moskewicz, Sergey Karayev, Ross B. Girshick, Trevor Darrell, Kurt Keutzer. 2014

439 citations recorded by the source corpus.

Convolutional Neural Networks (CNNs) can provide accurate object classification. They can be extended to perform object detection by iterating over dense or selected proposed object regions. However, the runtime of such detectors scales as the total number and/or area of regions to examine per image, and training such detectors may be prohibitively slow. However, for some CNN classifier topologies, it is possible to share significant work among overlapping regions to be classified. This paper presents DenseNet, an open source system that computes dense, multiscale features from the convolutional layers of a CNN based object classifier. Future work will involve training efficient object detectors with DenseNet feature descriptors.

Subfields: 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 Computer Vision and Pattern Recognition

In Model Architectures

Citations in the corpus

Measuring the Carbon Intensity of AI in Cloud Instances (2022, positive): “We trained three sizes of DenseNets [19] on MNIST [25].”
Deformable part models are convolutional neural networks (2014, neutral): “In the context of CNNs, this method dates back to (at least) early work on face detection in [31], and has been used in contemporary work including OverFeat [28], DetectorNet [3], and DenseNet [19].”
Speeding up Convolutional Neural Networks with Low Rank Expansions (2014, neutral): “To speed up test-time in a sliding window context for a CNN, [13] shows that multi-scale features can be computed efﬁciently by simply convolving the CNN across a ﬂattened multi-scale pyramid.”
SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <1MB model size (2016, neutral): “ImageNettrained DNNs have also been applied to a number of applications pertaining to autonomous driving, including pedestrian and vehicle detection in images [12, 7] and videos [2], as well as segmenting the shape of the road [1].”
DeepEdge: A multi-scale bifurcated deep network for top-down contour detection (2014, neutral): “Based on the runtimes reported in [14], if executed on a GPU our method would take about 5 minutes and could be made faster using the approach described in [12].”
Particular object retrieval with integral max-pooling of CNN activations (2015, positive): “…of the fully connected layers have been shown to be very effective and capable of adaptation to various domains (Oquab et al., 2014), such as scene recognition (Donahue et al., 2013; Sicre & Jurie, 2015), object detection (Iandola et al., 2014), and semantic segmentation (Girshick et al., 2014).”
Multi-scale Orderless Pooling of Deep Convolutional Activation Features (2014, neutral): “[12] mentioned earlier, and another is DenseNet [42].”
FireCaffe: Near-Linear Acceleration of Deep Neural Network Training on Compute Clusters (2015, neutral): “Accelerated DNN training solutions would solve a major pain point for these companies.”
Richer Convolutional Features for Edge Detection (2016, positive): “They employed DenseNet [27] to extract a feature vector for each pixel, and then SVM classier was used to classify each pixel into the edge or non-edge class.”
Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations (2021, positive): “We train a 121-layer DenseNet 43 , with weights initialized using ImageNet 42 .”
Comparing YOLOv3, YOLOv4 and YOLOv5 for Autonomous Landing Spot Detection in Faulty UAVs (2022, neutral): “The convolution architecture is based on modiﬁed DenseNet [33].”
Pedestrian detection with a Large-Field-Of-View deep network (2015, neutral): “Recent work has focused on reducing the computational time needed for detection of deep nets [34], [35].”
Intern VL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks (2023, neutral): “Starting with the pioneering AlexNet [55], a variety of convolutional neural networks (CNNs) have emerged, continuously refreshing the ImageNet benchmark [27, 32, 43, 47, 49, 72, 114, 124].”
ImageNet Large Scale Visual Recognition Challenge (2014, neutral): “3 - - Toyota Technological Institute at Chicago†, Ecole Centrale Paris‡ George Papandreou†, Iasonas Kokkinos‡(Papandreou, 2014; Papandreou et al., 2014; Jojic et al., 2003; Krizhevsky et al., 2012; Sermanet et al., 2013; Dubout and Fleuret, 2012; Iandola et al., 2014)”
IDRiD: Diabetic Retinopathy - Segmentation and Grading Challenge. (2020, positive): “4 ) in which they innovatively combined the DenseNets ( Iandola et al., 2014 ) and dilation block with U-Net ( Ronneberger et al., 2015 ) to capture more context information and multi-scale features.”
Deep Multi-patch Aggregation Network for Image Style, Aesthetics, and Quality Estimation (2015, neutral): “In [27, 20, 15], researchers computed ImageNet features (i.”
Fine-Tuning CNN Image Retrieval with No Human Annotation (2017, positive): “Specifically, using activations of CNNs trained for classification as the offthe-shelf image descriptors [4], [5] and adapting them for a number of tasks [6], [7], [8] have shown acceptable results.”
Real-Time Pedestrian Detection with Deep Network Cascades (2015, neutral): “Other researchers have focused specifically on speeding up Deep Neural Networks [19] [21] [23], but with no real-time solutions yet.”
Convolutional Channel Features (2015, neutral): “We also introduce techniques to accelerate the feature pyramid computation like feature maps approximation at nearby pyramid scales and shared convolution through patchwork [17] [28] [43], making CCF more computationally efficient under the sliding window setting.”