Loading...

Domain Adaptation Using Source Classifier for Object Detection

Mozafari, Azadeh Sadat | 2016

696 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 48888 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Jamzad, Mansour
  7. Abstract:
  8. Detection degradation caused by distribution discrepancy between the training and testing domains is a common problem in object detection systems. The difference between training and testing domains’ distribution mainly happenes because of the different ways of collecting and gathering data. For instance, datasets which have images with different illumination, view point, resolution, background and are obtained by different acquisition systems, have variance in distribution. The solution toward improving the detection rate of the classifier trained on training (source) domain when it is applied on testing (target) domain is to use Domain Adaptation (DA) techniques. One of important branches of DA methods is the model-transferring methods which use the trained source classifiers instead of source data for applying adaptation between two domains. The benefits of this branch of DA methods is their low computational (time and memory) complexity which makes them suitable for object detection systems. In this research, we focus on linear SVM classifiers and try to introduce two novel model-transferring DA methods for object detection problems. In the first proposed method, HMCA, obtaining higher detection rate and being applicable for heterogeneous source-target domains are the main achievements. According to the way of applying adaptation, it is also possible for HMCA to prpperly predict the amount of adaptability between two domains by measuring ISDP property between two domains. In the second proposed DA method, CA-SVM, the main concentration is on non-homogeneous distributed target domains. It seems when a target domain cannot be discriminated by a linear SVM classifier properly, dividing it to linearly separable subdomains and applying adaptation for locally trained linear classifiers can lead to higher adaptation rate. The advantage of CA-SVM in comparing to similar target division methods is in not using the source data to divide the target domain and instead use source classifier for applying division. Different experiments are conducted on human detection and image-classification datasets, comparing to several baselines. HMCA shows 0.02~0.07 in 10-3 FPPW improvement in detection and 2~4 percent improvement in classification. CA-SVM shows 0.05~0.1 in 10-3 FPPW improvement in detection and 3~16 percent improvement in classificataion, in comparing to considered baselines
  9. Keywords:
  10. Domain Adaptation ; Object Detection ; Target Domain ; Support Vector Machine (SVM)Classifier ; Transferring Domain Adaptation Method ; Images Classification

 Digital Object List

 Bookmark

No TOC