Incremental Boosting Incremental Learning of Boosted Face Detector
Incremental Boosting Incremental Learning of Boosted Face Detector ICCV 2007 Unsupervised Incremental Learning for Improved Object Detection in a Video CVPR 2012 2014. 7. 1. Jeany Son
Outline Supervised incremental learning for boosted classifier Incremental Learning of Boosted Face Detector, ICCV 2007 Unsupervised incremental learning for boosted classifier Unsupervised Incremental Learning for Improved Object Detection in a Video, CVPR 2012
Incremental Learning for Boosted Face Detector ICCV 2007
Face Detection - Hard examples
Incremental Learning Offline Learning vs. Incremental Learning - +
Domain-partitioned Real Adaboost is minimized when where
Loss function for incremental learning Domain-partitioning Strong classifier Likelihood of incremental learning: Linear combination of offline & online parts
Key issues 1) Adjustable parameters of the strong classifier H(x) 2) Estimation of Loff(H(x)) without offline samples 3) Choice of linear combination coefficient αy
Adjustable Parameters
F(x) is learned by means of some discriminative criterion e. g. KL divergence, Bhattacharyya distance Obtain proper domain partition of instance space for discrimination of different categories Small online samples to adjust F(x) is not unreasonable Offline training : determine F(x) Incremental training : adjust G(z)
Estimating Offline Loss function (without offline samples) Naïve Bayes
Online Loss function & Optimization Minimize loss using Steepest-decent method
Online reinforcement ratio of category y Contributions of online/offline sample
Datasets False alarm that is learned incrementally
Comparisons in CMU+MIT frontal face dataset False alarm that is learned incrementally
Unsupervised Incremental Learning for Boosted Detector CVPR 2012
Unsupervised Incremental MIL Crowded environment and cluttered background
Unsupervised Incremental MIL Contribution MIL based incremental learning for Real Adaboost Unsupervised online sample collection
Unsupervised Incremental MIL Offline detector Incremental Learning Negative samples False alarm Tracker Positive samples Missing or Low confidence Online sample collection
Online Sample Collection
Online Sample Collection Positive Samples : missing or low confidence detection Negative Samples : False alarm (30% overlap of detection with track response) Positive bag : consist of 10 patches around a missed detection Negative bag : one unmerged false alarm
MIL loss function Soft max/min soft loss function
Overfitting Avoidance Inc 1: base=offline detector Inc 0: base=previous iteration
Detection results False alarm Missing object
Detection results False alarm Missing object
- Slides: 30