Tata Consultancy Services Limited

Ramya Hebbalaguppe

Srinidhi Hegde

Puneet Gupta

Jitender Kumar Maurya

Ranjitha Prasad

Srinidhi Hedge

Implementations of the present disclosure build a Bayesian student network using the knowledge learnt by an accurate but complex pre-trained teacher network, and sparsity induced by the variational parameters in a student network. Further, the sparsity inducing capability of the teacher on the student network is learnt by employing a Block Sparse Regularizer on a concatenated tensor of teacher and student network weights. Specifically, the student network is trained using the variational lower bound based loss function, constrained on the hint from the teacher, and block-sparsity of weights.

Sparsity constraints and knowledge distillation based learning of sparser and compressed neural networks

Hand segmentation on wearable devices is a challenging computer vision problem with a complex background because of varying illumination conditions, computational capacity of device(s), different skin tone of users from varied race, and presence of skin color background. The present application provides systems and methods for performing, in real time, hand segmentation by pre-processing an input image to improve contrast and removing noise/artifacts. Multi Orientation Matched Filter (MOMF) is implemented and applied on the pre-processed image by rotating the MOMF at various orientations to form an edge image which comprises strong edges and weak edges. Weak edges are further removed using morphological operation. The edge image is then added to the input image (or pre-processed image) to separate different texture region in image. Largest skin-color blob is then extracted which is considered to be correct segmented hand.

Systems and methods for performing hand segmentation

Textual overlays/labels add contextual information in Augmented Reality (AR) applications. The spatial placement of labels is a challenging task particularly for real time videos. Embodiments of the present disclosure provide systems and methods for optimal placement of contextual information for Augmented Reality (AR) applications to overcome the limitations of occlusion with object/scene of interest through optimally placing labels aiding better interpretation of scene. This is achieved by combining saliency maps computed for each frame of an input video with Euclidean distance between current and previous overall positions for each frame based on an initial overlay position of the label to calculate an updated overlay position for label placement in the video. The placement of overlays is formulated as an objective function that minimizes visual saliency around the object of interest and minimizes the temporal jitter facilitating coherence in real-time AR applications.

Growing community of inventors

Gurgaon, India

Ramya Hebbalaguppe