Online and Active Learning for Pedestrian Detection
Jump to navigation
Jump to search
SVMs (Online and/or Structured Output)
- Fast Kernel Classifiers with Online and Active Learning. Bordes et al. JMLR 05. PDF (Online SVM that works well on INRIA.)
- Solving MultiClass Support Vector Machines with LaRank. Bordes et al. (ICML '07) PDF (Online Structured SVM)
- Sequence Labeling SVMs Trained in One Pass. Bordes et al. PDF (Online Structured SVM applied to text).
- A Support Vector Method for Multivariate Performance Measures. Thorsten Joachims. PDF (Structured SVMs can be used to optimize loss functions such as ROC area or precision/recall that might be more appropriate for the visual recognition task. In particular, my hypothesis is that handling these loss functions directly will allow us to do active learning more effectively across multiple points along the DET/ROC curve. My idea is that traditional margin based active learning is suboptimal because it samples along the margin specified by a single bias which is not necessarily related to the desired operating point. While 1e-4 FPPW seems to be the important operating point, the default SVM decision boundary is often located somewhere else, so margin based active learning is essentially sampling the wrong place). Actually, I take this back (hah). It would be hard to make this work in the online scenario, since the new loss functions don't decompose on a sample by sample basis. From the perspective of active learning, we are just trying to do as well as the passive learner using fewer labels. Since passive learning is learning the "wrong" decision boundary as well we don't need to worry about this here.
- Max-Margin Markov Networks. Taskar et al. NIPS 2004 pdf. Max-margin markov networks are a subset of structural SVMs, and were developed concurrently.
Active Learning (Machine Learning)
- Margin-based Active Learning for Structured Output Spaces. Dan Roth and Kevin Small, ECML '06. PDF
- Learning on the Border: Active Learning in Imbalanced Data Classification. Ertekin et al. PDF (This reference deals with imbalanced data and also appears to involve online scenarios.)
Active Learning (CS101.2 references)
- S. Dasgupta. "Analysis of a greedy active learning strategy" NIPS 2004 pdf
- bounds on number of labels required (bad news for general case)
- show simple greedy algorithm always achieves near optimal number of queries
- analysis sort of silly because so focused on finding realiazable hypothesis (0 training error)
- M. Balcan, S. Hanneke, J. Wortmann. "The true sample complexity of Active Learning", COLT 2008. pdf
- claim active learning almost always helps (contradicts almost all prev results)
- idea: easier to learn epsilon-good clf then to verify it is epsilon-good
- doesn't make sense, example 1.1 seems incorrect (let w=epsilon), poorly written paper
- W is problem dependent (depends on the true hypothesis, it's not a parameter of the learning algorithm.) Their statement is instead about the sample complexity as a function of 1/epsilon, which you can imagine as being a parameter under the "user's" control. Therefore it doesn't make sense to peg w=epsilon since w is not under our control, and epsilon is a free parameter.
- S. Dasgupta, D.J. Hsu. "Hierarchical sampling for active learning", ICML 2008. pdf
- R. Castro, C. Kalish, R. Nowak, R. Qian, T. Rogers, X. Zhu. "Human Active Learning", NIPS 2008. pdf
- R. Castro, R. Nowak. "Minimax Bounds for Active Learning", COLT 2007 pdf
- S. Dasgupta. "Coarse Sample complexity bounds for active learning", NIPS 2005. pdf
- S. Tong, D. Koller. "Support Vector Machine Active Learning with Applications to Text Classification." JMLR 2001 pdf
- S. Dasgupta, D. Hsu, C. Monteleoni. "A General Agnostic Active Learning Algorithm", NIPS 2007. pdf
Active Learning (Computer Vision)
- Automatically Labeling Data Using Multi-class Active Learning. Yan et al. ICCV03. PDF
- Active sampling via tracking. Roth et al. CVPRW08. [1]
- Active Learning for visual object detection (SEVILLE). Abramson and Freund. UCSD-TR05. PDF, pdf-v2?
- Active Learning in Face Recognition: Using Tracking to Build a Face Model. R. Hewitt & S. Belongie. CVPR06. pdf
- Entropy-based active learning for object recognition. A Holub, P Perona, MC Burl. CVPRW08. pdf
- Active learning with gaussian processes for object categorization. Kapoor, Grauman, Urtasun, and Darrell. ICCV07. pdf
- Efficient Boosting-Based Active Learning for Specific Object Detection Problems. Nguyen, Binh, Bischof. 2008. pdf
- Practical online active learning for classification. Monteleoni and Kaariainen. CVPR07 pdf
- Active learning based pedestrian detection in real scenes. Yang, Li, Pan, Zhao, Zhu. ICPR06. pdf
- Two-Dimensional Active Learning for image classification. Qi, ..., Zhang. CVPR08. pdf
- Automatically labeling video data using multi-class active learning. Yan, Yang, Hauptmann. ICCV03. pdf