Worth Reading

PAMI

Volume 30, Issue 1

Groups of Adjacent Contour Segments for Object Detection

1. Introduction. Many existing recognition algorithms use local patches (color cue) instead of contour features. This paper proposes a scale-invariant local shape features formed by chains of k connected roughly straight contour segments.

2. The shape represented by k-AS increases with k, as 2-AS can be L shapes, 3-AS can form C, F, and Z shapes.

3. Following a bag-of-features paradigm, they construct a codebook of k-AS types

4. Related Work. Problem to existing solutions: clutter and restricted class of shapes. A natural thought is that people can use boosting to select edge patterns for classification (not scale invariant and tailored to specific class, dataset).

5. Split the detector window into separated tilts and compute histograms in each tilt so as to form a descriptor.

6. Use of Integral Histograms for speeding up the sliding-window detector based on kAS histogram.

7. Conclusion. Future work, Multiple View. Rotation invariant. More complex framework.

Locally Rotation, Contrast and Scale Invariant Descriptors for Texture Analysis

Yet another feature descriptor.

Mixture of Spherical Distributions for Single-View Relighting

1. Assumptions. Given a 3D model of the object, the method can recover the direction and intensity of ultiple light sources and the number of light sources and specular reflectance property of the object. Previous methods assume 1) all light sources are infinitely distance (directional) 2) the geometry of the target object is known, 3) number of light sources known.

2. To estimate the parameters for the light sources, the paper use EM algorithm.

LEGClust -A clustering Algorithm Based on Layered Entropic Subgraphs

1. This clustering algorithm uses local structure information of the data, so it can be used to cluster Swiss Roll like data.

Mutual Information for Lucas-Kanade Tracking (MILK): An Inverse Compositional Formulation.

Classical LKT algorithm uses SSD as the distance measure for two images, this paper proposes Mutual Information as a different measure. And it also propose an efficient algorithm based on Back Composition. Source code available.