Abstract: For many real-world problems, the training set does not well represent the class separability information present in the test samples necessary for classification. This is particularly true when the training and the test samples are obtained from different environments, thus producing quite dissimilar feature distributions. Domain adaptation (DA) is a specific type of “transfer learning” (TL), where one can use abundant training samples available in source domain to aid a statistical learning task on test samples present in the target domain. This talk will introduce the problems of DA and TL, followed by brief discussions of popular and generic methods of solution. This will be followed by two of our recent approaches of DA applied in classification of objects in images. One will use Eigen-analysis of the data samples and the other will be based on minimization of an optimization function formulated using landmark points on manifolds. The final part of the talk will discuss our very recent approach of using DA over a deep-CNN, termed Transfer-CNN or Deep-DA, which will aid in the task of Face Recognition in degraded scenarios as in surveillance conditions.
Abstract: Digitally preserving archaeological artifacts demand 3D modeling of those artifacts which in turn require high-quality 3D data acquired by a 3D scanner. In this talk, I shall present the design perspectives of a low-cost portable 3D laser scanning mechanism designed and fabricated by us. It is fully automatic, non-invasive, and able to generate high-quality data sets. We propose this design for its wide applicability to the digitization of archaeological artifacts, 3D modeling, automated defect inspection, object recognition, virtual reality simulation etc. The scanner consists of a CCD camera in conjunction with a line laser projector. The laser projector is mounted on a rotating scan head assembly. The 3D surface information is captured using laser plane triangulation. The scanner is able to acquire the surface shape data in the form of dense 3D point cloud. Due to self-occlusion and concavities in the scanned object, it is possible to scan only a part of the object from one point of view. Therefore, we need to perform multiple scans and combine scan data from multiple wide baseline viewpoints onto a common reference coordinate system. Moreover, due to the unstructured environment, the pose and the position of the scans are not known beforehand and need to be generated from the information within the scan data itself. 3D mesh data acquisition is often afflicted by undesirable measurement noise. Such noise has an aversive impact to further processing and also to human perception and hence plays a pivotal role in mesh processing. We present here a fast saliency-based algorithm that can reduce the noise while preserving the finer details of the original object. The proposed algorithm finds wide application in digitization of archaeological artifacts, such as statues and sculptures, where it is of paramount importance to capture the 3D surface with all its details as accurately as possible.
Abstract: In bad weather conditions like fog and haze, the particles present in the atmosphere (e.g. dust particles, water droplets) scatter incident light in different directions. As a result, images taken under these condition suffer from reduced visibility and lack of contrast. Image dehazing method tries to recover a haze-free portrayal of the given hazy image. In this talk we discuss about two different approaches. First we discuss a method that can dehaze an image independent f whether it was captured during the day or night. The second method is based on the fundamental tool of human perception: comparison.
Abstract: Advances in technology have made the use of wearable cameras affordable and practical. The cameras are often harnessed on wearer's head, chest or shoulder and provide a unique opportunity to capture the world from the first person perspective of the camera wearer. While, the ability to share one's own point of view made such egocentric cameras popular in consumer domain, the cameras have quickly gained popularity in sports, law enforcement, life logging, assistive vision, as well as virtual and augmented reality. The long, redundant, and extremely shaky nature of egocentric videos make them hard to view from start to end, thereby necessitating use of automated computer vision tools for their efficient consumption. In this talk I will describe some of the works done by my group in this area such as temporal segmentation, action recognition, summarization, as well as visual SLAM for egocentric videos.
Abstract: Facebook, Apple, Google, and Samsung products have popularized Virtual Reality (VR) and Haptics. In general, Human Machine Interfaces (HMI) are rapidly evolving with information processing, communication and robotics. VR and Haptic interfaces are a relatively recent HMI that enable the human user to touch, feel, and manipulate either virtual objects or real objects at a distance through teleportation, intrinsically a multidisciplinary research area. In this talk I will present some of our research works on Medical Simulation using indigenously developed cutting-edge haptic technologies in the Touchlab at IIT Madras and our startup company Merkel Haptics.