About
62
Publications
32,090
Reads
How we measure 'reads'
A 'read' is counted each time someone views a publication summary (such as the title, abstract, and list of authors), clicks on a figure, or views or downloads the full-text. Learn more
2,806
Citations
Introduction
Researcher in deep learning, video processing, time series analysis, applied machine learning.
Additional affiliations
November 2018 - present
Navya
Position
- Technical Lead Machine Learning
Description
- - Deep learning for perception tasks in autonomous driving (AD) : Object detection, semantic segmentation, multi-task learning, semantic segmentation on pointclouds, Simulation-to-reality domain transfer, Real-time production deployment of deep learning models - Advising master internship, & phd students in reinforcement learning, pointcloud processing - Deep learning for autonomous driving Workshop Principal co-organizer. - Deep reinforcement learning for AD global overview & knowledge
October 2016 - August 2017
Position
- Lecturer
Description
- Machine learning Dimensionality reduction Reccomendation systems Informatics
May 2017 - December 2017
Uncanny Vision Systems
Position
- Consultant
Description
- Focus : Video representation learning for anomaly detection Models : Autoencoders (Convolutional, Contractive, Spatio-temporal, C3D), Predictive models(ConvLSTM, Linear predictors), Generative models(GANs, VAEs)
Publications
Publications (62)
Semantic Bird's Eye View (BEV) maps offer a rich representation with strong occlusion reasoning for various decision making tasks in autonomous driving. However, most BEV mapping approaches employ a fully supervised learning paradigm that relies on large amounts of human-annotated BEV ground truth data. In this work, we address this limitation by p...
Active Learning (AL) has remained relatively unexplored for LiDAR perception tasks in autonomous driving datasets. In this study we evaluate Bayesian active learning methods applied to the task of dataset distillation or core subset selection (subset with near equivalent performance as full dataset). We also study the effect of application of data...
Autonomous driving (AD) perception today relies heavily on deep learning based architectures requiring large scale annotated datasets with their associated costs for curation and annotation. The 3D semantic data are useful for core perception tasks such as obstacle detection and ego-vehicle localization. We propose a new dataset, Navya 3D Segmentat...
Active Learning (AL) has remained relatively unexplored for LiDAR perception tasks in autonomous driving datasets. In this study we evaluate Bayesian active learning methods applied to the task of dataset distillation or core subset selection (subset with near equivalent performance as full dataset). We also study the effect of application of data...
Autonomous driving (AD) perception today relies heavily on deep learning based architectures requiring large scale annotated datasets with their associated costs for curation and annotation. The 3D semantic data are useful for core perception tasks such as obstacle detection and ego-vehicle localization. We propose a new dataset, Navya 3D Segmentat...
Modern object detection architectures are moving towards employing self-supervised learning (SSL) to improve performance detection with related pretext tasks. Pretext tasks for monocular 3D object detection have not yet been explored yet in literature. The paper studies the application of established self-supervised bounding box recycling by labeli...
Annotating objects with 3D bounding boxes in LiDAR pointclouds is a costly human driven process in an autonomous driving perception system. In this paper, we present a method to semi-automatically annotate real-world pointclouds collected by deployment vehicles using simulated data. We train a 3D object detector model on labeled simulated data from...
Autonomous driving (AD) datasets have progressively grown in size in the past few years to enable better deep representation learning. Active learning (AL) has re-gained attention recently to address reduction of annotation costs and dataset size. AL has remained relatively unexplored for AD datasets, especially on point cloud data from LiDARs. Thi...
Modern object detection architectures are moving towards employing self-supervised learning (SSL) to improve performance detection with related pretext tasks. Pretext tasks for monocular 3D object detection have not yet been explored yet in literature. The paper studies the application of established self-supervised bounding box recycling by labeli...
Data augmentation is a key component of CNN based image recognition tasks like object detection. However, it is relatively less explored for 3D object detection. Many standard 2D object detection data augmentation techniques do not extend to 3D box. Extension of these data augmentations for 3D object detection requires adaptation of the 3D geometry...
With the development of deep representation learning, the domain of reinforcement learning (RL) has become a powerful learning framework now capable of learning complex policies in high dimensional environments. This review summarises deep reinforcement learning (DRL) algorithms and provides a taxonomy of automated driving tasks where (D)RL methods...
Background-Foreground classification is a well-studied problem in computer vision. Due to the pixel-wise nature of modeling and processing in the algorithm, it is usually difficult to satisfy real-time constraints. There is a trade-off between the speed (because of model complexity) and accuracy. Inspired by the rejection cascade of Viola-Jones cla...
Point cloud datasets for perception tasks in the context of autonomous driving often rely on high resolution 64-layer Light Detection and Ranging (LIDAR) scanners. They are expensive to deploy on real-world autonomous driving sensor architectures which usually employ 16/32 layer LIDARs. We evaluate the effect of subsampling image based representati...
Point cloud datasets for perception tasks in the context of autonomous driving often rely on high resolution 64-layer Light Detection and Ranging (LIDAR) scanners. They are expensive to deploy on real-world autonomous driving sensor architectures which usually employ 16/32 layer LIDARs. We evaluate the effect of subsampling image based representati...
With the development of deep representation learning, the domain of reinforcement learning (RL) has become a powerful learning framework now capable of learning complex policies in high dimensional environments. This review summarises deep reinforcement learning (DRL) algorithms, provides a taxonomy of automated driving tasks where (D)RL methods ha...
The use of hyperspectral imaging for medical applications is becoming more common in recent years. One of the main obstacles that researchers find when developing hyperspectral algorithms for medical applications is the lack of specific, publicly available, hyperspectral medical data. The work described in this paper was developed within the framew...
Lidar has become an essential sensor for autonomous driving as it provides reliable depth estimation. Lidar is also the primary sensor used in building 3D maps which can be used even in the case of low-cost systems which do not use Lidar. Computation on Lidar point clouds is intensive as it requires processing of millions of points per second. Addi...
Deep Reinforcement Learning (DRL) has become increasingly powerful in recent years, with notable achievements such as Deepmind's AlphaGo. It has been successfully deployed in commercial vehicles like Mobileye's path planning system. However, a vast majority of work on DRL is focused on toy examples in controlled synthetic car simulator environments...
Consider a family $Z=\{\boldsymbol{x_{i}},y_{i}$,$1\leq i\leq N\}$ of $N$ pairs of vectors $\boldsymbol{x_{i}} \in \mathbb{R}^d$ and scalars $y_{i}$ that we aim to predict for a new sample vector $\mathbf{x}_0$. Kriging models $y$ as a sum of a deterministic function $m$, a drift which depends on the point $\boldsymbol{x}$, and a random function $z...
Lidar has become an essential sensor for autonomous driving as it provides reliable depth estimation. Lidar is also the primary sensor used in building 3D maps which can be used even in the case of low-cost systems which do not use Lidar. Computation on Lidar point clouds is intensive as it requires processing of millions of points per second. Addi...
Lidar has become an essential sensor for autonomous driving as it provides reliable depth estimation. Lidar is also the primary sensor used in building 3D maps which can be used even in the case of low-cost systems which do not use Lidar. Computation on Lidar point clouds is intensive as it requires processing of millions of points per second. Addi...
Surgery for brain cancer is a major problem in neurosurgery. The diffuse infiltration into the surrounding normal brain by these tumors makes their accurate identification by the naked eye difficult. Since surgery is the common treatment for brain cancer, an accurate radical resection of the tumor leads to improved survival rates for patients. Howe...
Confusion matrix results of the SVM supervised classification with polynomial kernel applying the 10-fold cross validation method to each patient.
(DOCX)
Confusion matrix results of the SVM supervised classification with linear kernel applying the 10-fold cross validation method to each patient.
(DOCX)
Confusion matrix results of the SVM supervised classification with RBF kernel applying the 10-fold cross validation method to each patient.
(DOCX)
Confusion matrix results of the SVM supervised classification with sigmoid kernel applying the 10-fold cross validation method to each patient.
(DOCX)
Hyperspectral imaging (HSI) allows for the acquisition of large numbers of spectral bands throughout the electromagnetic spectrum (within and beyond the visual range) with respect to the surface of scenes captured by sensors. Using this information and a set of complex classification algorithms, it is possible to determine which material or substan...
Videos represent the primary source of information for surveillance applications and are available in large amounts but in most cases contain little or no annotation for supervised learning. This article reviews the state-of-the-art deep learning based methods for video anomaly detection and categorizes them based on the type of model and criteria...
In the class of streaming anomaly detection algorithms for univariate time series, the size of the sliding window over which various statistics are calculated is an important parameter. To address the anomalous variation in the scale of the pseudo-periodicity of time series, we define a streaming multi-scale anomaly score with a streaming PCA over...
Background-Foreground classification is a fundamental well-studied problem in computer vision. Due to the pixel-wise nature of modeling and processing in the algorithm, it is usually difficult to satisfy real-time constraints. There is a trade-off between the speed (because of model complexity) and accuracy. Inspired by the rejection cascade of Vio...
Random forests perform bootstrap-aggregation by sampling the training samples with replacement. This enables the evaluation of out-of-bag error which serves as a internal cross-validation mechanism. Our motivation lies in using the unsampled training samples to improve each decision tree in the ensemble. We study the effect of using the out-of-bag...
![Fig. 2: HOP vs BOP: Ultrametric contour map (UCM [3]), hierarchy (top)...](profile/Bangalore-Kiran/publication/275885699/figure/fig1/AS:294631020482562@1447256917379/HOP-vs-BOP-Ultrametric-contour-map-UCM-3-hierarchy-top-and-a-braid-of-partitions_Q320.jpg)
In obtaining a tractable solution to the problem of extracting a minimal partition from hierarchy or tree by dynamic programming, we introduce the braids of partition and h-increasing energies, the former extending the solution space from a hierarchy to a larger set, the latter describing the family of energies, for which one can obtain the solutio...
This theoretical paper provides a basis for the optimality of scale-sets by Guigues [6] and the optimal pruning of binary partition trees by Salembier-Garrido [11]. They extract constrained-optimal cuts from a hierarchy of partitions. Firstly, this paper extends their results to a larger family of partitions, namely the braid [9]. Secondly, the pap...
Cette étude concerne le partitionnement d'un ensemble de telle sorte que les séparations entre classes soient matérialisées. On le résoud, dans les cas continu et discret, au moyen de hiérarchies de tesselations dont les classes sont des ouverts réguliers. Dans le cas discret, le passage partition→tessellation s'exprime par des topologies d'Alexand...
This theoretical paper introduces a novel continuous representation of hierarchy of partitions, and generalizes the conditions of h-increasingness and scale increasingness [1] to obtain a global-local optimum on the hierarchy. It studies in particular the Lagrange optimization problem and gives the condition on the energy to achieve constrained opt...
Hierarchical segmentation has been a model which both identifies with the construct of extracting a tree structured model of the image, while also interpreting it as an optimization problem of the optimal scale selection. Hierarchical processing is an emerging field of problems in computer vision and hyper-spectral image processing community, on ac...
This paper introduces a new fusion paradigm that reorders the contours of an input hierarchy of segmentations utilizing any external function. Ground truth partitions, markers, numerical functions, and other additional sources are used to obtain a new reordered hierarchy. This transformation aids in combining sets, functions, and partitions in diff...
In the current technical note we provide a topological generalization of hierarchy of partitions(HOP) structure, and the implications over the axioms of h-increasingness and scale increasingness [13]. Further in this study we will explicit the Lagrange optimization in the optimal cuts problem and the conditions necessary on the energy to obtain a g...
A hierarchy of segmentations(partitions) is a multiscale set representation of the image. This paper introduces a new set of scale space operators or transformations on the space of hierarchies of partitions. An ordering of hierarchies is proposed which is endowed by an ω-ordering based on a global energy over the classes of the hierarchy. A class...
In evaluating a hierarchy of segmentations H of an image by ground truth G, which can be partitions of the space or sets, we look for the optimal partition in H that “fits” G best. Two energies on partial partitions express the proximity from H to G, and G to H. They derive from a local version of the Hausdorff distance. Then the problem amounts to...
A new approach is proposed for finding optimal cuts in hierarchies of partitions by energy minimization. It rests on the notion of h-increasingness, allows to find best(optimal) cuts in one pass, and to obtain nice “climbing” scale space operators. The ways to construct h-increasing energies, and to combine them are studied, and illustrated by two...
Hierarchical segmentation is a multi-scale analysis of an image and provides a series of simplifying nested partitions. Such a hierarchy is rarely an end by itself and requires external criteria or heuristics to solve problems of image segmentation, texture extraction and semantic image labelling. In this theoretical paper we introduce a novel fram...
The paper deals with global constraints for hierarchical segmentations. The proposed framework associates, with an input image, a hierarchy of segmentations and an energy, and the subsequent optimization problem. It is the first paper that compiles the different global constraints and unifies them as Climbing energies. The transition from global op...
A new approach is proposed for finding the ”best cut” in a hierarchy of partitions by energy minimization. Said energy must be ”climbing” i.e. it must be hierarchically and scale increasing. It encompasses separable energies [5], [9] and those which composed under supremum [14], [12]. It opens the door to multivariate data processing by providing l...
A new approach is proposed for finding the "best cut" in a hierarchy of
partitions by energy minimization. Said energy must be "climbing" i.e. it must
be hierarchically and scale increasing. It encompasses separable energies and
those composed under supremum.
Summarization of cricket videos is very important because of three reasons: 1) its long duration making manual highlights generation tedious 2) less explored area compared to other sports like soccer 3) huge viewership. We propose a novel summarization scheme for cricket which exploits its contextual semantics. First, we detect the bowling frames b...
With increasing computational requirements for state-of-art computer vision and graphics algorithms, multi-core processor implementations have picked up momentum. Adapting single core image processing algorithms for multi-core environment is difficult. Different platforms require a hand tailored approach to optimally fit the algorithm into a given...
This paper describes a novel approach to the connected component labeling problem, derived from two fast labeling algorithms, Wu et al. and Park et al. We propose a method that improves over existing divide and conquer methods. We propose two new methods - First, hierarchical (coarse to fine) label propagation from various sub images. Second, the r...
With increasing computational requirements for state-of-art computer vision and graphics
algorithms, multi-core processor implementations have picked up momentum. Adapting single
core image processing algorithms for multi-core environment is difficult. Different platforms
require a hand tailored approach to optimally fit the algorithm into a given...
Questions
Question (1)
The 3D-DLAD (3D Deep Learning for Autonomous Driving) workshop is organized as part of the flagship automotive conference, Intelligent Vehicles http://iv2019.org.
The previous edition of this workshop, Deep learning for autonomous driving (DLAD) was held at ITSC 2017, Japan.
The presentations/talks from previous edition of the workshop can be found here https://sites.google.com/site/dlitsc17/program
For this edition, we are soliciting contributions in the domain of deep learning for 3D data applied to autonomous driving in (but not limited to) the following topics.
- Deep Learning for Lidar based object detection and/or tracking
- Deep Learning for Lidar point-cloud clustering and road segmentation
- Deep Learning for computer vision point-cloud processing (VSLAM, meshing, inpainting)
- Deep Learning for TOF sensor based driver monitoring
- Deep Learning for Odometry and Map/HDmaps generation with Lidar cues
- Deep fusion of automotive sensors (Lidar, Camera, Radar)
- Design of datasets (Synthetic Lidar sensors & Transfer learning)
- Cross-modal feature extraction for Sparse output sensors like Lidar
- Generalization techniques for different Lidar sensors, multi-Lidar setup and point densities
- Lidar based maps, HDmaps, prior maps, occupancy grids
- Real-time implementation on embedded platforms (Efficient design & hardware accelerators)
- Challenges of deployment in a commercial system (Functional safety & High accuracy)
- New lidar based technologies and sensors
- End to end learning of driving with Lidar information (Single model & modular end-to-end)
- Deep learning for dense Lidar point cloud generation from sparse Lidars and other modalities
Workshop : https://sites.google.com/view/3d-dlad-iv2019/ Location : Paris, France
Submission : 7th Feb 2019 (submission portal is not open yet)
Acceptance Notification : 29th March 2019 Workshop Date : 9th June 2019
Please feel free to contact us if there are any questions. We are sorry if this is a repost.
Abstract: Deep Learning has become a de-facto tool in Computer Vision and 3D processing with boosted performance and accuracy for diverse tasks such as object classification, detection, optical flow estimation, motion segmentation, mapping, etc. Lidar sensors are playing an important role in the development of Autonomous Vehicles as they overcome some of the many drawbacks of a camera based system, such as degraded performance under changes in illumination and weather conditions. In addition, Lidar sensors capture a wider field of view, and directly obtain 3D information. This is essential to assure the security of the different agents and obstacles in the scene. It is a computationally challenging task to process more than 100k points per scan in realtime within modern perception pipelines. Following the said motivations, finally to address the growing interest on deep representation learning for lidar point-clouds, in both academic as well as industrial research domains for autonomous driving, we invite submissions to the current workshop to disseminate the latest research.
