Universitat Politècnica de Catalunya (UPC), Barcelona Faculty of Informatics (FIB) Kernel-based Machine Learning and Multivariate Modelling (MAI-KMLMM) Final Course Project
This is an implementation for the kernelized KISSME algorithm in R as described by Nguyen and De Baets (2019) for large-scale metric learning from equivalence constraints. The algorithm has been applied to the VIPeR dataset by Gray et al., a python implementation of the LOMO feature extractor (original paper by Liao et al.) is used for preprocessing the data. For more detail, please have a look into the final report.
- The run.py script in the LOMO Feature Extractor folder needs to be run in order to create the features
- Start the test.R in the main directory. It automatically
- imports the feature vectors
- splits up the dataset into training, validation and test
- genereates must-link and cannot-link constraints on the training dataset
- evaluates different hyperparameter combinations (kernel methods and epsilon) on the validation set
- returns test scores
- visualizes the results
- Nguyen and De Baets, "Kernel Distance Metric Learning Using Pairwise Constraints for Person Re-Identification," IEEE Transactions on Image Processing, Vol. 28, 2019
- D. Gray, S. Brennan, and H. Tao, "Evaluating Appearance Models for Recognition, Reacquisition, and Tracking,", IEEE International Workshop on Performance Evaluation for Tracking and Surveillance (PETS), 2007
- Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z. Li, "Person Re-identification by Local Maximal Occurrence Representation and Metric Learning," CVPR2015