Exact Acceleration of Linear Object Detectors

By Charles Dubout, Franccois Fleuret. 2012

155 citations recorded by the source corpus.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition, 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 Aerospace Engineering

In Computer Vision and Pattern Recognition

In Efficiency & Compression

Citations in the corpus

ImageNet Large Scale Visual Recognition Challenge (2014, background): “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)”
Deformable part models are convolutional neural networks (2014, neutral): “, [7, 24]).”
Voting for Voting in Online Point Cloud Object Detection (2015, methodology): “It is well known that a linear classifier in the case of sliding window detection is equivalent to convolution, and therefore standard techniques such as the Fast Fourier Transform (FFT) can be applied to compute the detection scores efficiently [3].”
DenseNet: Implementing Efficient ConvNet Descriptor Pyramids (2014, positive): “Using the Bottom-Left Fill (BLF) algorithm implemented in [21], we stitch the multiple scales of the input image pyramid onto as many large (often of size 1200x1200 or 2000x2000, depending on available GPU memory) images as needed, and then run each individual image as a batch.”
Convolutional Channel Features (2015, methodology): “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.”
Deformable Part Models with CNN Features (2014, methodology): “To achieve efficiency during convolutions with the part templates (used both during training and testing), we perform convolutions using the Fast Fourier Transform, along the lines of [2].”
The Fastest Deformable Part Model for Object Detection (2014, methodology): “It has been found in previous works [10, 22, 8] that in DPM most of the time is spent on calculating the appearance term due to the high dimension.”
A survey on face detection in the wild: Past, present and future (2015, unknown): “Wewilldescribe severalapproachesforreducingthecomplexitywithoutsignificantperformancedegradation[165] [166] [167] [168] [169].”
Learning coarse-to-fine sparselets for efficient object detection and scene classification (2015, methodology): “Meanwhile, it should be pointed out that the speedup factors are obtained by only using our sparselets framework, higher speedup factors can be easily obtained by combining our method with other existing accelerating algorithms, such as Cascade [45], FFT [46], branch-and-bound [47], coarse-to-fine [”
A coarse-to-fine approach for fast deformable object detection (2011, methodology): “Dubout et al. [29] show how to speed-up the convolution of the object model with HOG features using the Fourier Transform.”
30Hz Object Detection with DPM V5 (2014, methodology): “FFLD [11] 0.”
Tracking Revisited Using RGBD Camera: Unified Benchmark and Baselines (2013, methodology): “There are many potential ways to speed up the algorithm, such as using [9] to accelerate the naive convolution for sliding window detector, or use an GPU version of optical flow [24] to replace the CPU version [5].”
Robust Multi-resolution Pedestrian Detection in Traffic Scenes (2013, methodology): “Specially for our implementation, we modified the code of the FFT based implementation [15] for the fast convolution computation.”
Learning Everything about Anything: Webly-Supervised Visual Concept Learning (2014, methodology): “This testing step can be easily sped up by leveraging recent fast detection methods [10, 14, 46].”
Training Deformable Part Models with Decorrelated Features (2013, unknown): “, [8]), making further acceleration possible.”
Efficient Processing of Deep Neural Networks: A Tutorial and Survey (2017, methodology): “However, there are several drawbacks to using FFT: (1) the benefits of FFTs decrease with filter size; (2) the size of the FFT is dictated by the output feature map size which is often much larger than the filter; (3) the coefficients in the frequency domain are complex.”
Modeling local and global deformations in Deep Learning: Epitomic convolution, Multiple Instance Learning, and sliding window detection (2015, methodology): “For this, we first draw inspiration from the ‘image patchwork’ technique introduced in [9] and exploited in DCNNs by [12].”
Deformable Part Descriptors for Fine-Grained Recognition and Attribute Prediction (2013, background): “We have explored two options for dealing with this: one based on a per-component appearance representation with no semantic parts; the other leverages additional annotations to estimate semantic correspondence of the latent parts across components.”
Object Detection in 20 Years: A Survey (2019, methodology): “mputational complexity. Fig. 16. An illustration of how to speed up a linear detector (e.g., HOG detector, DPM, etc) in frequency domain with fast Fourier transform and inverse fast Fourier transform [226]. convolution operation, and is the point-wise product. The above calculation can be accelerate”
Volumetric Semantic Segmentation Using Pyramid Context Features (2013, methodology): “See the supplementary material for a demonstration of the improvement in efficiency yielded by using “pyramid filtering” instead of pre-existing techniques, such as slidingwindow [12] or FFT-based filtering [13].”