THE PRELIMINARY ORAL EXAMINATION FOR THE DEGREE OF Ph.D. IN COMPUTER SCIENCE FOR
Fatemeh Mir Rashed
Building visual models of objects robust to "extrinsic" variations such as camera view angle, resolution, lighting, and blur has long been one of the challenges in computer vision. Generally, a discriminative or generative statistical model is learned by acquiring a large set of examples, extracting low-level features which encode shape, color, or texture from the segmented or cropped objects, and finally, training the model (usually a classifier) using the extracted features vectors.
Applied to a test image, however, the trained model usually works if the training set was representative of the test set, i.e., if the distribution over training examples roughly matches the distribution of the test data. Unfortunately, there are often cases when this implicit key assumption of learning algorithms is violated, resulting in a sharp performance drop. There has been a recent growing interest in the machine learning community to develop effective mechanisms to transfer or adapt knowledge from one (source) domain to another related (target) domain. While these advances have also been recently applied in visual domains with promising results, object models are still being trained and tested on images consisting of only one object zoomed and cropped at the center of a relatively uniform background. As a result, in such experimental settings the general problem of object recognition is reduced to that of image classification.
We present a framework for adaptive object detection using Transfer Component Analysis (TCA), an unsupervised domain adaptation and dimensionality reduction technique. Given labeled examples from the source domain and unlabeled examples from the target domain, we obtain a transformation to a latent subspace that reduces the distance between the source and target distributions while simultaneously preserving data properties, enabling standard classifiers to generalize directly to unseen examples from the target domain. Unlike recent domain adaptation work in computer vision, which generally focuses on image classification, we address the problem of extreme class imbalance present when performing domain adaptation for object detection. We apply our technique to vehicle detection in a challenging urban surveillance data set, demonstrating the performance of our approach with various amounts of supervision, including the fully unsupervised case.
The proposed future work includes extending the current framework to multiple source domains, multiple object categories, and using class labels from the target domains when they are available.
Examining Committee:
Dr. Larry S. Davis - Chair
Dr. Amitabh Varshney - Dept’s Representative
Dr. David Jacobs - Committee Member
EVERYBODY IS INVITED TO ATTEND THE PRESENTATION