Arjun Sharma

Hi, I am Arjun Sharma. I am a research engineer at Zoox, Amazon's self-driving car subsidiary. I obtained my MS in Robotics at the Robotics Institute, Carnegie Mellon University where I worked with Prof. Kris Kitani. Previously, I was a research engineer at Intrinsic, an Alphabet company working on industrial robotics and a researcher in the Multimedia group at Xerox Research Centre in Bangalore, India. I graduated with a bachelor's degree in Mathematics and Computing from Indian Institute of Technology (IIT) Guwahati in 2014.

My research interests are in the fields of Reinforcement Learning, Deep Learning and Computer Vision. At XRCI, I worked on finding solutions using Deep Learning to variety of problems in computer vision such as object recognition and detection in RGB-D images, activity classification in videos, text spotting in handwritten and printed document images and focus estimation. At CMU, I worked on Deep Reinforcement Learning and Inverse Reinforcement Learning.

Publications

DirectedInfo-GAIL: Learning Hierarchical Policies from Unsegmented Demonstrations using Directed Information. Arjun Sharma, Mohit Sharma, Nicholas Rhinehart, Kris Kitani. ICLR 2019

The use of imitation learning to learn a single policy for a complex task that has multiple modes or hierarchical structure can be challenging. In fact, previous work has shown that when the modes are known, learning separate policies for each mode or sub-task can greatly improve the performance of imitation learning. In this work, we discover the interaction between sub-tasks from their resulting state-action trajectory sequences using a directed graphical model. We propose a new algorithm based on the generative adversarial imitation learning framework which automatically learns sub-task policies from unsegmented demonstrations. Our approach maximizes the directed information flow in the graphical model between sub-task latent variables and their generated trajectories. We also show how our approach connects with the existing Options framework, which is commonly used to learn hierarchical policies. pdf video

Phase-Parametric Policies for Reinforcement Learning in Cyclic Environments. Arjun Sharma, Kris Kitani. AAAI 2018

In many reinforcement learning problems, parameters of the MDP may vary with its phase while the agent attempts to learn through its interaction with the environment. Attempting to model such phase variations with a standard data-driven approach (e.g. deep networks) without explicitly modeling the phase of the model can be challenging. Ambiguities may arise as the optimal action for a given state can vary depending on the phase. To better model cyclic environments, we propose phase-parameterized policies and value function approximators that explicitly enforce a cyclic structure to the policy or value space. pdf video code

GeThR-Net: A Generalized Temporally Hybrid Recurrent Neural Network for Multimodal Information Fusion. Ankit Gandhi, Arjun Sharma, Arijit Biswas, Om Deshmukh. ECCV 2016 workshop on Computer Vision for Audio-Visual Media

In this paper, we propose a novel generalized deep neural network architecture where temporal streams from multiple modalities can be combined.There are total M+1 (M is the number of modalities) components in the proposed network. The first component is a novel temporally hybrid RNN that exploits the complimentary nature of the multimodal temporal information by allowing the network to learn both modality specific temporal dynamics as well as the dynamics in a multimodal feature space. M additional components are added to the network which extract discriminative but non-temporal cues from each modality. Finally, the predictions from all of these components are linearly combined using a set of automatically learned weights. We perform exhaustive experiments on three different datasets spanning four modalities. The proposed network is relatively 3.5%, 5.7% and 2% better than the best performing temporal multimodal baseline for UCF-101, CCV and Multimodal Gesture datasets respectively. pdf arxiv code

LIVELINET: A Multimodal Deep Recurrent Neural Network to Predict Liveliness in Educational Videos. Arjun Sharma, Arijit Biswas, Ankit Gandhi, Sonal Patil, Om Deshmukh. EDM 2016, Best paper award nominee

Online educational videos have emerged as one of the most popular modes of learning in the recent years. Studies have shown that liveliness is highly correlated to engagement in educational videos. While previous work has focused on feature engineering to estimate liveliness and that too using only the acoustic information, in this paper we propose a technique based on the combination of audio and visual features to predict liveliness. We also propose a novel multimodal deep recurrent neural network based approach to automatically estimate if an educational video is lively or not. On the StyleX dataset of 450 one-minute long educational video snippets, our approach shows an absolute improvement of 8.9% in the classification accuracy over the baseline. pdf

Enhancing RGB CNNs with Depth. Arjun Sharma, Pramod Sankar K. ACPR 2015

Most current approaches for recognition in RGB-D images fall in either the late fusion or the early fusion category. A drawback of the early fusion scheme is its inapplicability when one of the modalities is absent at test time. On the other hand, a late fusion of features does not allow the correlated nature of modalities to be exploited effectively. Recent approaches using Deep Learning are not immune to these problems either. In this project, we proposed a simple, yet elegant method towards combining early and late fusion of colour and depth information when training Deep Convolutional Neural Networks. We show that when fine-tuning CNNs, an intermediate depth pre-training step provides a significant jump in colour recognition accuracy. pdf ppt

Adapting Off-the-Shelf CNNs for Word Spotting and Recognition. Arjun Sharma, Pramod Sankar K. ICDAR 2015

The word spotting approach is extremely useful for searching and annotating documents for which robust recognizers are unavailable. Traditionally, hand-designed features were used to represent the word images for spotting. In this paper, we learn a data-driven representation for word-images from Convolutional Neural Networks (CNNs). Previous approaches that learn deep neural networks for a particular task/dataset are difficult to design and train for generic word spotting. Instead, by “adapting” a CNN trained for a different problem, we show tremendous speedup in the training phase. Our experiments show that features extracted from an adapted-CNN handsomely outperform hand-designed features on both spotting and recognition tasks for printed (English and Telugu) and handwritten (IAM) document collections. pdf

Emotion recognition in the wild. ICMI 2013

The goal of the Emotion Recognition in the Wild Challenge 2013 was to develop a system to classify emotions expressed by the primary human subject in the short video clips of acted scenes lasting approximately one-two seconds, including the audio track which may contain human voices as well as background music.

I was a member of Université de Montréal's team that participated in this challenge. I worked on adapting Wang et al's activity recognition using dense trajectories pipeline for emotion recognition. I was also involved in generating a video-emotion label dataset from descriptive video service files of movies. pdf

Other Projects

Object Grasping using Hierarchical Reinforcement Learning

Course project: Mechanics of Manipulation

In this project, we developed a hierarchical reinforcement learning based method for object grasping as a part of a course project, where object grasping is considered as consisting of two sub-tasks - reaching and grasping. Hierarchical Reinforcement Learning is used to learn two policies - a high-level manager policy and a low-level policy. Knowledge sharing between subtasks is improved by shifting from traditional hard-switching to soft switching of subtasks, where the manager policy learns to appropriately weigh the reward due to various sub-goals at each step. pdf

Improving Semantic Segmentation using Inverse Reinforcement Learning

Course project: Computer Vision

In this work, we look at the segmentation of aerial images into road, building and background classes. We improve the ability of a CNN trained on this task to identify roads using IRL. Trajectories of vehicles moving on the road serve as the ‘expert demonstrations’ for IRL. We aim to complement appearance cues learnt by the CNN with an important piece of information about affordance of roads – that vehicles move on them. ppt

Focus estimation in images

Arjun Sharma, Pramod Sankar K

In this project, we used a pre-trained Deep Convolutional Neural Network to identify pixels in an image which are in focus. We trained a network on a dataset of wildlife images collected from the web and showed superior performance as compared to other recent methods. We then used this network to find focused regions in videos of house inspection recorded using drones and found the performance on this different domain to be surprisingly good.

Evaluation of Multi-class SVMs using DEA

Guide: Prof. V. Vijaya Saradhi

One-vs-One, One-vs-All, and Error correcting output codes (ECOC) are some of the most widely employed methods for multi-class classification using binary classifiers. The superiority of any one method over the others has been the subject of much research. In this project, we compared these methods on the basis of parameters that describe an optimal separating hyperplane in the best possible way, namely the training error, the S-span, the fraction of support vectors and the margin of individual hyperplanes unlike the traditional method of comparing them along the test error. We proposed a novel method of using Data Envelopment Analysis (DEA) to compute the relative efficiencies of the hyperplanes instead of comparing the entire ensemble as one. pdf

Developing a Machine Translation System using Translation Memory

Guide: Dr. Sandipan Dandapat

The aim of the project was to improve upon present Translation memory systems. Given an input sentence, the search for the closest match in the translation memory in the source language was improved using an IR engine and tree based methods as compared to the traditional edit distance based approaches. The difference in the source sentence, its closest match and the translation in the memory was identified and automatic post-editing carried out using statistical machine translation to obtain an accurate translation for the source sentence. ppt1 ppt2

Activity Recognition in Egocentric Videos using Deep Learning

Vardaan Pahuja, Arjun Sharma, Ankit Gandhi, Arijit Biswas

In this project, we compared the performance of state-of-the-art deep learning methods against previous techniques on the activities of daily living (ADL) dataset of first person videos. We compared R-CNN object detectors with previously used DPM based object detectors. We also compared the performance of Long-Short Memory (LSTM) networks with the previously used temporal pyramid based approach to aggregate per frame object detection scores for video classification.