Papers by Abhijit Sharang
2015 IEEE Winter Conference on Applications of Computer Vision, 2015
Lecture Notes in Computer Science, 2014
Automatic understanding and modelling deviant behaviour is a challenging task, especially when pe... more Automatic understanding and modelling deviant behaviour is a challenging task, especially when performed in unsupervised manner. Most techniques today focus either on trajectory clustering or capturing intrinsic scene features to detect and identify the abnormal content in videos. In this paper, we model the usual and dominant behaviour of videos using unsupervised probabilistic topic models, as compliment of which we identify the "anomalous" ones. We present the features to define finite set of visual words, which are then projected to latent space using models like pLSA. The abnormal events in the videos are then marked owing to their low-likelihood of occurrence. We further describe the generalization of the existing model to context-independent scenario using quantized spatio-temporal features and generative models, and discuss potential ideas for localization of anomaly. Experiments and results have been shown on surveillance video dataset.
Modelling top down visual attention is a challenging task owing to the huge dimensionality of the... more Modelling top down visual attention is a challenging task owing to the huge dimensionality of the input involved.Moreover,since it is task specific,the variables to be taken into consideration might differ with every task.This project aims at developing a general model for top down attention and testing the effectiveness of the model on two different kinds of tasks.Results show that the model works well when the person is dynamic with respect to the environment,but does not perform upto the expectations when the person is static.
Segmenting similar objects from multiple images is a challenging problem due to variation in imag... more Segmenting similar objects from multiple images is a challenging problem due to variation in image properties.In this project,we have examined the effectivenss of saliency as a feature for segmenting objects from multiple images.We divide the images into various regions and then attempt to merge similar regions in order to obtain a region which contains the most salient object.Our experiments show that saliency is more effective than other image features for co-segmentation
iitk.ac.in
Determining the authenticity of an image is now an important area of research .In our work,we att... more Determining the authenticity of an image is now an important area of research .In our work,we attempt to classify whether a digital image is a genuine image or is a manipulated version of some authentic image.We have applied three different algorithms discussed in [1] and [2] with modifications of our own. We have then implemented two techniques discussed in [5] and [6] to detect the fake regions in the image. Our approach is invariant to the type of manipulation that has been done to the images.From our methods,we are able to classify nearly 80% manipulated images.
Talks by Abhijit Sharang
• Co-segmentation aims to segment common objects from a collection of images given by the user.
Drafts by Abhijit Sharang
Language dictionaries are a high quality linguistic resource curated by trained professionals and... more Language dictionaries are a high quality linguistic resource curated by trained professionals and are key to human learning of language. In our work, we attempt to use the definitions in dictionaries for learning word embeddings by mapping the definitions to their target word embeddings through an RNN-LSTM model. We experiment with two versions of the output embedding : in the first, we keep the output embedding representing the term fixed to pre-trained GloVe word vectors, and in the second, we jointly learn the input em-beddings which represent the definitions and the output embeddings which represent the term. We proceed to evaluate the word vectors on various tasks defined and used by existing literature for word embedding evaluation. Our word vectors trained on dictionaries perform significantly better on certain important tasks such as distinguishing similarity from re-latedness.
Uploads
Papers by Abhijit Sharang
Talks by Abhijit Sharang
Drafts by Abhijit Sharang