Research
From MIPAL
Contents
Feature Selection and Extraction
In many machine learning and pattern recognition problems, the number of attributes are too large that the learning process becomes complicated with degraded performance. To alleviate this problem, depending on the problems to be solved, one may select good attributes from the available input attributes or extract new features by combining the existing attributes. The former is known as input feature selection while the latter is called as feature extraction.
We have tackled the feature selection and extraction problems extensively and proposed the following algorithms.
- Feature selection for classification problems
- Feature extraction for classification problems
- Feature extraction for regression problems
- Robust feature extraction using non-L2 norm
- PCA-L1 (Principal Component Analysis based on L1-norm maximization)
- a short tutorial is also available
- PCA-Lp (Principal Component Analysis based on L1-norm maximization)
- LDA-Lp (Linear Discriminant Analysis by Lp-norm)
- L1BDA (L1-norm based Biased Discriminant Analysis)
- BDAGM (Biased Discriminant Analysis using Generalized Mean)
- PCA-L1 (Principal Component Analysis based on L1-norm maximization)
- Theory in kernel methods
- NPT (Nonlinear Projection Trick)
- a short tutorial is also available
- NPT (Nonlinear Projection Trick)
Object Recognition
Object Detection
- Rainbow SSD "Enhancement of SSD by concatenating feature maps for object detection", BMVC2017, London, UK, Sep. 2017. (Code available)
- RUN "Two-Layer Residual Feature Fusion for Object Detection", 8th International Conference on Pattern Recognition Applications and Methods (ICPRAM2019), Prague, Czech Republic, February 2019. (arXiv)
- MMOD "Mixture-Model-based Bounding Box Density Estimation for Object Detection", arXiv, Nov. 2019.