Vijay Kumar
Identifying ragas in Indian Music
Raga is the central structure of Indian classical music characterized by an appropximately repeating sequences or subseqences of notes. Automatically recognizing Raga from a short audio segment has several interesting applications in digital music indexing, recommendation and retrieval. We attempted to identify the ragas by capturing two important characteristic features of Indian music - pakads and gamakas. Specifically, we solve the raga classification problem in a non-linear SVM framework using a kernel formed by the combination of two kernels to capture raga similarities in pakads and gamakas space. (in submission..)
Annotating actors in Videos
We consider the problem of automatic annotation of faces in videos such as movies, given a large database of labeled faces. This can have applications in video indexing, content based search, surveillance, and real time recognition on wearable computers. We formulated the problem as a semi-supervised transductive learning that propagates the labels from a seed set to unlabeled faces by incorporating the constraints in temporal and feature space. The seed set for propagation is obtained by labeling the video frames using sparse representation framework using l1 minimization and selecting only confident labelings. (in submission..)
Face Recognition with Limited Labeled Samples
Existing state-of-the-art still image face recognition algorithm based on sparse representation framework assumes the availability of a large number of labeled training examples. However, in many practical problems, such as photo-tagging, surveillance, etc, the number of labeled training samples are very limited leading to significant degradations in classification performance. We extend the current state-of-the-art and exploit the large number of unlabeled samples available in practical problems mentioned above. We propose a graph based semi-supervised algorithm that labels the unlabeled samples through a multi-stage label propagation combined with sparse representation. (pdf)
Indian Movie Face Database
We created a large database of Indian movie actors to facilitate the face recognition research in unconstrained settings such as videos. The database consists of 34,512 face images of 100 Indian actors collected from approximately 103 video clips and movies. It includes 67 male and 33 female actors with at least 200 images for each actor. We chose movies as the faces in movies exhibit large variety of variations in resolution, illumination, age, makeup, pose, occlusion, viewpoints, and expression. Movies are selected carefully in such a way that they produce large age variations for actors. (pdf)
Document Image Restoration
We introduce a sparse representation based framework to solve the inverse problem of restoring degraded document images. We make an observation that different characters in a language possess similar strokes, curves, and edges, and thus learn a dictionary that captures these similarities. (pdf)
Emergency alert system using GSM and GPS
The project was aimed at making a portable system which will help people in remote locations send out an SOS for help, by combining the GSM and GPS technologies. The system which consists of a micro-controller, a Tx-Rx circuit, a GSM modem and a GPS receiver is activated when the person in need of help, presses the button from a hand-held device with transmitter, thereby sending a signal to the emergency system. Micro-controller after recieving the signal communicates with the GPS system to get the exact location of the person, and sends out a pre-recorded message with all the details to the nearest authority or stored contacts over a GSM network. (pdf)
