ISI Digital Commons (Indian Statistical Institute )
Not a member yet
7571 research outputs found
Sort by
Diving into the Depths of Spotting Text in Multi-Domain Noisy Scenes
No full textWhen used in a real-world noisy environment, the capacity to generalize to multiple domains is essential for any autonomous scene text spotting system. However, existing state-of-the-art methods employ pretraining and fine-tuning strategies on natural scene datasets, which do not exploit the feature interaction across other complex domains. In this work, we explore and investigate the problem of domain-agnostic scene text spotting, i.e., training a model on multi-domain source data such that it can directly generalize to target domains rather than being specialized for a specific domain or scenario. In this regard, we present the community a text spotting validation benchmark called Under-Water Text (UWT) for noisy underwater scenes to establish an important case study. Moreover, we also design an efficient super-resolution based end-to-end transformer baseline called DA-TextSpotter which achieves comparable or superior performance over existing text spotting architectures for both regular and arbitrary-shaped scene text spotting benchmarks in terms of both accuracy and model efficiency. The dataset, code and pre-trained models have been released in our Github
DRL-based Multi-Stream Scheduling of Inference Pipelines on Edge Devices
No full textReal-time scheduling of multiple neural network-based inference pipelines on Graphics Processing Unit (GPU) based edge devices is an active area of research nowadays. Applications like Advanced Driver-Assistance Systems (ADAS) execute multiple such inference pipelines to make informed decisions on driving scenarios. The real-time performance of ADAS is often limited by platform resource limitations and thus incurs execution latency which ultimately leads to deadline violation. In this regard, modern GPUs provide support for concurrent execution of multiple compute streams. However, there is a lack of scheduling strategies in the literature that consider multiple such compute streams and focus on the concurrent execution of inference pipelines for more efficient real-time scheduling. In this paper, we address this issue by proposing a Deep Reinforcement Learning (DRL) based solution for multi-stream scheduling of inference pipelines on edge GPUs. Using DRL, we learn how to map every layer of the target inference pipelines to high or low-priority streams, while satisfying task-level deadline requirements. The experimental evaluation shows the efficacy of the proposed approach as compared to some baseline approaches
Experimental Investigation and Its Parametric Analysis of AWJM Process Parameters on Surface Quality of Green Composites
No full textThe present work aims to investigate the effects of the abrasive water jet machining (AWJM) process on the surface quality of green composites (GC) through an experimental investigation and parametric analysis. Experiments are conducted-based Taguchi (L27) OA by varying AMGS, SoD, NS, AMFR & WP, and the corresponding responses like surface quality (Ra) are measured. Thereafter, effect of process parameters and most optimal combinations for AWJM process is identified through parametric and ANOVA. Furthermore, utilizing liner regression analysis, a mathematical model is created for the best Ra prediction for the AWJM procedure. In order to validate with the experimental findings, the confirmatory analysis has finally been tested. According to the results, Ra is significantly influenced by the parameters, namely AMGS, AMFR, SoD, and NS. The ideal setting for AWJM process is obtained as AMFR as 2 g/s, WP as 80MPa, SoD as 2.5 mm, NS as 200 mm/min, and AMGS as 40 mesh
On Fourier Analysis of Sparse Boolean Functions over Certain Abelian Groups
No full textGiven an Abelian group G, a Boolean-valued function f : G → {−1, +1}, is said to be s-sparse, if it has at most s-many non-zero Fourier coefficients over the domain G. In a seminal paper, Gopalan et al. [15] proved “Granularity” for Fourier coefficients of Boolean valued functions over ℤn2 , that have found many diverse applications in theoretical computer science and combinatorics. They also studied structural results for Boolean functions over ℤn2 which are approximately Fourier-sparse. In this work, we obtain structural results for approximately Fourier-sparse Boolean valued functions over Abelian groups G of the form, G := ℤn1p1 × · · · × ℤntpt, for distinct primes pi. We also obtain a lower bound of the form 1/(m2s)⌈φ(m)/2⌉, on the absolute value of the smallest non-zero Fourier coefficient of an s-sparse function, where m = p1 · · · pt, and φ(m) = (p1 − 1) · · · (pt − 1). We carefully apply probabilistic techniques from [15], to obtain our structural results, and use some non-trivial results from algebraic number theory to get the lower bound. We construct a family of at most s-sparse Boolean functions over ℤnp , where p \u3e 2, for arbitrarily large enough s, where the minimum non-zero Fourier coefficient is o(1/s). The “Granularity” result of Gopalan et al. implies that the absolute values of non-zero Fourier coefficients of any s-sparse Boolean valued function over ℤn2 are Ω (1/s). So, our result shows that one cannot expect such a lower bound for general Abelian groups. Using our new structural results on the Fourier coefficients of sparse functions, we design an efficient sparsity testing algorithm for Boolean function, which tests whether the given function is s-sparse, or ϵ-far from any sparse Boolean function, and it requires poly((ms)φ(m), 1/ϵ)-many queries. Further, we generalize the notion of degree of a Boolean function over an Abelian group G. We use it to prove an Ω(√s) lower bound on the query complexity of any adaptive sparsity testing algorithm
REAL-NET: A Monochromatic Depth Estimation Using REgional Attention and Local Feature Mapping
No full textEstimating the depth of a scene from a color image is an arduous task as it requires prior depth information to remove any uncertainty in the 3D interpretation. In recent works, many supervised depth estimation methods gave promising results, learning the priors based on end-to-end training. Unfortunately, inadequate addressing of actual physical constraints leads to inaccurate estimation. In this paper, we present REAL-Net, a supervised depth estimation encoder-decoder net by incorporating REgional constraints in the encoder through the Attention mechanism, and neighborhood feature constraints in the decoder taking care of the Local mapping. These make the output significantly clear and robust to the structural and color information. Results obtained upon conducting extensive experiments on benchmark datasets namely KITTI, Foggy Zurich, and Cityscapes establish the proposed method’s capability to outperform results obtained with the existing approaches
New hybrid gene selection-sample classification method in microarray data
No full textThe gene expression dataset generated by DNA microarray technology contains expression profiles of huge quantities of genes for very small samples. Among these genes, a very small number of genes are informative for cancer sample identification and classification. Informative genes finding is an essential task of microarray gene expression data analysis. Here, a new hybrid gene selection-sample classification model (NHGSSC) is proposed for selection of relevant genes and classification of cancer samples. The NHGSSC performs two tasks-gene selection and sample classification. For gene selection, a new hybrid single filter and α-depth limited best first search based single wrapper method (SFα-BFSSW) is proposed. From these subsets, highly informative genes are selected by counting frequency of occurrence (FO) of every gene. Then SFα-BFSSW method-based ensemble classifier (SFα-BFSSWEC) is built by combining the classifiers created for the selected gene subsets. Experimental results demonstrate the superiority of the NHGSSC to other existing models
Gathering at Fixed Nodes by Oblivious Mobile Robots
No full textA swarm of robots is a distributed system of small, autonomous, inexpensive mobile robots that work collaboratively to achieve some specific goal. This thesis aims to study the gathering problem, which is a fundamental coordination problem. In the initial configuration, the robots are located at distinct nodes of an input graph. The main objective of the problem is that all the robots must meet at a point. In this thesis, the deployment space of the robots is assumed to be an infinite grid graph. This thesis considers the gathering over meeting nodes problem in an infinite grid do- main. The input grid graph consists of some nodes which are marked as meeting nodes. The problem requires the mobile robots to gather at one of the pre-fixed meeting nodes. Several variations of the problem have been considered. All the initial configurations for which the problem is deterministically unsolvable have been characterized. A deter- ministic distributed algorithm has been proposed to solve the problem for the remaining configurations. The problem has also been studied under the objective function that minimizes the total number of moves made by all the robots to finalize the gathering task. A deterministic distributed algorithm has been proposed to solve the problem for all those solvable configurations, and the initial configurations for which the problem is unsolvable have been characterized. The proposed gathering algorithm is optimal with respect to the total number of moves performed by all the robots in order to finalize the gathering. Another variation of the problem considers two disjoint and finite groups of anonymous robots. The initial configurations also consist of two finite and disjoint sets of heteroge- neous meeting nodes. The objective is to design a distributed algorithm that gathers all the robots belonging to the first team at one of the meeting nodes belonging to the first type. Similarly, all the robots in the second team must gather at one of the meeting nodes belonging to the second type. The initial configurations for which the gathering problem is unsolvable have been characterized. For the remaining initial configurations, a distributed gathering algorithm has been proposed. The parking problem can also be considered as a variation of the problem. As a solution to the parking problem, the robots need to partition themselves into groups so that each parking node contains a number of robots equal to the capacity of the node. It is assumed that the number of robots in the initial configuration represents the sum of the capacities of the parking nodes. Under this setup, we have characterized all the initial configurations and the values of ki for which the problem is unsolvable. For the remaining configurations, a deterministic distributed algorithm was proposed
On Combinatorial Constructions of Mutually Unbiased Bases (MUBs) with Approximations
No full textConstruction of Mutually Unbiased Bases (MUBs) is a very challenging combinatorial problem in the domain of quantum information theory with several long-standing open questions. The basic definition considers two orthonormal bases in the d-dimensional complex Hilbert space with a fixed specific inner product value among any two vectors from two different bases, which is one by square root of d. Similarly, some k orthonormal bases are called Mutually Unbiased Bases (MUBs) if they are pairwise Mutually Unbiased. Reaching the upper bound on this k is believed to be an extremely challenging problem for more than half a century. Because of the difficulties in constructing a significantly large number of MUBs, the problem is relaxed, and the concept of Approximate Mutually Unbiased Bases (AMUBs) has been introduced. Here, the inner product of two vectors drawn from two different bases is upper bounded by some value instead of being exactly the same, namely one by the square root of d. In this thesis, we present different construction methods of Approximate MUBs using combinatorial structures such as Hadamard matrices and Resolvable Block Designs. We also propose the concept of Almost Perfect MUBs (APMUB), where we restrict the absolute value of the inner product into two values, one of which is zero. Among several constructions, we also exploit Mutually Orthogonal Latin Squares (MOLS) in this regard. Our results have implications for several existing combinatorial objects, such as Bi-angular vectors. With the understanding of APMUBs, we revisit a larger class of AMUBs and improve the results further in terms of larger classes for composites that are not prime powers and for both real and complex. The technique is more generalized in terms of exploring novel instances of RBDs that provide improved results. Finally, a heuristic framework is presented to search for approximate MUBs with significantly good parameters, and experimental outcomes of the computer programs are studied. Our proposed idea considers two parts. First, we apply basis reduction techniques (that are well-studied in Machine Learning literature) in obtaining the initial solutions. Then, we exploit the steepest ascent kind of search to improve the results further. The efficacy of our technique is shown through the construction of AMUBs in dimensions d = 6, 10, 46. From a more generic view, this approach considers approximately solving a challenging (where efficient deterministic algorithms are not known) mathematical problem in a discrete domain through state-of-the-art heuristic ideas. To summarize, in this thesis, we exploit several involved combinatorial techniques in a disciplined manner and also propose a heuristic to construct approximate MUBs
A deep learning based cognitive model to probe the relation between psychophysics and electrophysiology of flicker stimulus
No full textThe flicker stimulus is a visual stimulus of intermittent illumination. A flicker stimulus can appear flickering or steady to a human subject, depending on the physical parameters associated with the stimulus. When the flickering light appears steady, flicker fusion is said to have occurred. This work aims to bridge the gap between the psychophysics of flicker fusion and the electrophysiology associated with flicker stimulus through a Deep Learning based computational model of flicker perception. Convolutional Recurrent Neural Networks (CRNNs) were trained with psychophysics data of flicker stimulus obtained from a human subject. We claim that many of the reported features of electrophysiology of the flicker stimulus, including the presence of fundamentals and harmonics of the stimulus, can be explained as the result of a temporal convolution operation on the flicker stimulus. We further show that the convolution layer output of a CRNN trained with psychophysics data is more responsive to specific frequencies as in human EEG response to flicker, and the convolution layer of a trained CRNN can give a nearly sinusoidal output for 10 hertz flicker stimulus as reported for some human subjects
A note on convex solutions to an equation on open intervals
No full textThe note is concerned with the functional equation (Formula presented.) which is a generalised form of the so-called polynomial-like iterative equation. We investigate the existence of nondecreasing convex (both usual and higher order) solutions to this equation on open intervals using the Schauder fixed point theorem. The results supplement those proved by Trif (Aquat Math, 79:315–327, 2010) for the polynomial-like iterative equation by generalising them to a greater extent. This assertion is supported by some examples illustrating their applicability