Representation of local geometry in the visual system

By Jan J. Koenderink, A. Doom. 1987

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Neuroscience

Citations in the corpus

Hypercolumns for object segmentation and fine-grained localization (2014, background): “Koenderink and van Doorn [26] represented each pixel using “jets”, which are sets of partial derivatives of intensity up to a particular order, and described how edge orientation, curvature etc. can be estimated using this representation.”
Inside-Outside Net: Detecting Objects in Context with Skip Pooling and Recurrent Neural Networks (2015, methodology): “The first uses a multiscale representation [20] that captures fine-grained details by pooling from multiple lower-level convolutional layers in a ConvNet [37].”
Deformable Convolutional Networks (2017, background): “y the ﬁlter weights are learned, not the sampling locations like ours. This work is applied for video and stereo prediction. Combination of low level ﬁlters Gaussian ﬁlters and its smooth derivatives [30] are widely used to extract low level image structures such as corners, edges, T-junctions, etc.”
A Biologically Inspired System for Action Recognition (2007, methodology): “Space-time-oriented S1 units: Most studies focus on the spatial structure of receptive fields: V1 simple cells’ receptive field profiles were modeled by twodimensional Gabors or Gaussian-derivative functions in [28, 29].”
Structured Receptive Fields in CNNs (2016, methodology): “For i) we choose the family of Gaussian filters and its smooth derivatives for which it has been proven [12] that 3-rd or 4-th order is sufficient to capture all local image variation perceivable by humans.”
A comprehensive review of current local features for computer vision (2008, methodology): “Existing descriptors can be categorized into the following ﬁve types: (i) ﬁlter-based descriptors [17,25,61], (ii) distribution-based descriptors [9,34,37,42,48], (iii) textons [40,81,82], (iv) derivative-based descriptors [23,36,63], and (v) others [12,79,80].”
A sparse texture representation using local affine regions (2005, background): “features like differential invariants [18], [43] and filter”
Fully Convolutional Networks for Semantic Segmentation (2017, background): “van Doorn [27], we call our feature hierarchy the deep jet.”
Contour and Texture Analysis for Image Segmentation (2001, background): “We recommend Koenderink and van Doorn (1987, 1988) , Jones and Malik (1992), and Malik and Perona (1992) for a discussion of other advantages of such a representation.”
Automatic detection and recognition of signs from natural scenes (2004, unknown): “We call this vector a jet[10], which is used to represent the position of local features.”
Dynamic Steerable Blocks in Deep Residual Networks (2017, methodology): “In practice for CNNs working on images we investigate four typical choices of frames: i) the vanilla orthogonal pixel basis, ii) Gaussian derivatives, one of the most widely used overcomplete frames from the computer vision literature (also used in SIFT) [13, 23, 26, 30], iii) Framelets, a non-ortho”
PCA-SIFT: a more distinctive representation for local image descriptors (2004, background): “Mikolajczyk and Schmid [14] presented a comparative study of several local descriptors including steerable filters [4], differential invariants [9], moment invariants [18], complex filters [16], SIFT [11], and cross-correlation of different types of interest points [6, 13].”
Behavior recognition via sparse spatio-temporal features (2005, background): “[26] used differential descriptors [16] for their spatio-temporal interest points, however, among the descriptors examined for 2D features, differential descriptors are not particularly robust.”
Matching Widely Separated Views Based on Affine Invariant Regions (2004, methodology): “Lindeberg and Gȧrding (1997) on the other hand have developed a method to find blob-like regions using an iterative scheme, in the context of shape from texture.”
On Space-Time Interest Points (2005, unknown): “Using derivatives, we define event descriptors from the third order local jet 4 ( Koenderink and van Doorn, 1987 ) computed at spatio-temporal scales determined from the detection scales of the corresponding interest points”