Deep roto-translation scattering for object classification

By Edouard Oyallon, Stéphane Mallat. 2014

240 citations recorded by the source corpus.

Dictionary learning algorithms or supervised deep convolution networks have considerably improved the efficiency of predefined feature representations such as SIFT. We introduce a deep scattering convolution network, with complex wavelet filters over spatial and angular variables. This representation brings an important improvement to results previously obtained with predefined features over object image databases such as Caltech and CIFAR. The resulting accuracy is comparable to results obtained with unsupervised deep learning and dictionary based representations. This shows that refining image representations by using geometric priors is a promising direction to improve image classification and its understanding.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

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 Computer Vision and Pattern Recognition

Citations in the corpus

Scattering Networks for Hybrid Representation Learning (2018, methodology): “A scattering transform with scale J can be interpreted as a CNN with depth J [41], whose channels indexes correspond to different scattering frequency indexes, which is a structuration.”
Differentially Private Learning Needs Better Features (or Much More Data) (2020, methodology): “In this paper, we propose to use the ScatterNet features of Oyallon & Mallat (2015) as a basis for shallow differentially private vision classifiers.”
Stacked Capsule Autoencoders (2019, background): “Geometrical Reasoning Other attempts at incorporating geometrical knowledge into neural networks include exploiting equivariance properties of group transformations (Cohen and Welling, 2016) or new types of convolutional ﬁlters (Oyallon and Mallat, 2015; Dieleman et al., 2016).”
General E(2)-Equivariant Steerable CNNs (2019, methodology): “As a result, we are able to implement a wide range of equivariant models, covering regular GCNNs [6, 7, 8, 9, 10, 11], classical Steerable CNNs [1], Harmonic Networks [12], gated Harmonic Networks [2], Vector Field Networks [13], Scattering Transforms [14, 15, 16, 17, 18] and entirely new architectu”
Group Equivariant Convolutional Networks (2016, methodology): “Scattering networks have been extended to use convolutions on the group of translations, rotations and scalings, and have been applied to object and texture recognition (Sifre & Mallat, 2013; Oyallon & Mallat, 2015).”
Gauge Equivariant Convolutional Networks and the Icosahedral CNN (2019, background): “…Hartford et al., 2018), graphs (Kondor et al., 2018b; Hy et al., 2018; Maron et al., 2019), translations and rotations of the plane and 3D space (Oyallon & Mallat, 2015; Cohen & Welling, 2016; 2017; Marcos et al., 2017; Weiler et al., 2018b;a; Worrall et al., 2017; Worrall & Brostow, 2018;…”
RotDCF: Decomposition of Convolutional Filters for Rotation-Equivariant Deep Networks (2018, background): “The joint convolution across space and rotation has been studied in the scattering transform [25, 33].”
Harmonic Networks: Deep Translation and Rotation Equivariance (2016, methodology): “[24] use a larger number of rotations for texture classification and [26] also use many rotated handcrafted filter copies, opting not to learn the filters.”
Understanding deep convolutional networks (2016, background): “A roto-translation scattering [29, 27] linearizes them with wavelet ﬁlters along translations and rotations, with G j = SE ( n ) for all j > 1.”
On Lazy Training in Differentiable Programming (2018, neutral): “In this second set of experiments, we investigate whether variations of the models trained above in a lazy regime could increase the performance and, in particular, could outperform other linear methods which also do not involve learning a representation [26, 25].”
AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transformations Rather Than Data (2019, methodology): “…including rotations, translations, scaling and even more such as manipulating It is worth pointing out on CIFAR-10, the other reported methods [22, 7, 26, 21, 10] are usually based on different unsupervised networks and supervised classiﬁers for evaluation, making it difﬁcult to make a direct…”
Scaling the Scattering Transform: Deep Hybrid Networks (2017, neutral): “With appropriate discretization [27], { A J x, W 1 x } is ap-proximatively an isometry on the set of signals with limited bandwidth, and this implies the energy of the signal is pre-served.”
A General Theory of Equivariant CNNs on Homogeneous Spaces (2018, neutral): “Other related work includes [24–31].”
CubeNet: Equivariance to 3D Rotation and Translation (2018, background): “Sifre & Mallat [8] and Oyallon & Mallat [9] use a scattering network [35] for roto-translation invariant classification.”
Building a Regular Decision Boundary with Deep Networks (2017, background): “Geometric variabilities can be successfully handled by prior representations such as scattering transforms [23, 4, 20, 19, 25].”
Multiscale Hierarchical Convolutional Networks (2017, methodology): “The network takes as an input a color image x ( u, v 0 ) , with three color channels 1 ≤ v 0 ≤ 3 and u = ( u 1 , u 2 ) .”
A Mathematical Theory of Deep Convolutional Neural Networks for Feature Extraction (2015, background): “A nother aspect that was found to be desirable in practice [20], [33], but is not contained i n Mallat’s theory [19], is sub-sampling to reduce redundancy in the extracted features.”