Akshay Asthana

ORCID: 0000-0001-6871-346X
Publications
Citations
Views
---
Saved
---
About
Contact & Profiles
Research Areas
  • Face recognition and analysis
  • Face and Expression Recognition
  • Video Surveillance and Tracking Methods
  • Advanced Image and Video Retrieval Techniques
  • Emotion and Mood Recognition
  • Domain Adaptation and Few-Shot Learning
  • Advanced Neural Network Applications
  • Sentiment Analysis and Opinion Mining
  • Medical Imaging and Analysis
  • Anomaly Detection Techniques and Applications
  • 3D Shape Modeling and Analysis
  • Natural Language Processing Techniques
  • Image Retrieval and Classification Techniques
  • Biometric Identification and Security
  • Human Pose and Action Recognition
  • Digital Mental Health Interventions
  • Mental Health Research Topics
  • Multimodal Machine Learning Applications
  • Fuzzy and Soft Set Theory
  • Gaze Tracking and Assistive Technology
  • Medical Image Segmentation Techniques
  • Video Analysis and Summarization
  • Image Processing and 3D Reconstruction
  • Text and Document Classification Technologies
  • Speech and dialogue systems

Imperial College London
2013-2014

Australian National University
2009-2011

Mitsubishi Electric (United States)
2011

Jaypee Institute of Information Technology
2007

We present a novel discriminative regression based approach for the Constrained Local Models (CLMs) framework, referred to as Discriminative Response Map Fitting (DRMF) method, which shows impressive performance in generic face fitting scenario. The motivation behind this is that, unlike holistic texture features used AAM approaches, response map can be represented by small set of parameters and these very efficiently reconstructing unseen maps. Furthermore, we show that adopting simple...

10.1109/cvpr.2013.442 article EN 2009 IEEE Conference on Computer Vision and Pattern Recognition 2013-06-01

The development of facial databases with an abundance annotated data captured under unconstrained 'in-the-wild' conditions have made discriminative deformable models the de facto choice for generic landmark localization. Even though very good performance localization has been shown by many recently proposed techniques, when it comes to applications that require excellent accuracy, such as behaviour analysis and motion capture, semi-automatic person-specific or even tedious manual tracking is...

10.1109/cvpr.2014.240 article EN 2009 IEEE Conference on Computer Vision and Pattern Recognition 2014-06-01

An ideal approach to the problem of pose-invariant face recognition would handle continuous pose variations, not be database specific, and achieve high accuracy without any manual intervention. Most existing approaches fail match one or more these goals. In this paper, we present a fully automatic system for that only meets requirements but also outperforms other comparable methods. We propose 3D normalization method is completely leverages accurate 2D facial feature points found by system....

10.1109/iccv.2011.6126336 article EN International Conference on Computer Vision 2011-11-01

We propose a method for automatic emotion recognition as part of the FERA 2011 competition. The system extracts pyramid histogram gradients (PHOG) and local phase quantisation (LPQ) features encoding shape appearance information. For selecting key frames, K-means clustering is applied to normalised vectors derived from constraint model (CLM) based face tracking on image sequences. Shape closest cluster centers are then used extract features. demonstrate results SSPNET GEMEP-FERA dataset. It...

10.1109/fg.2011.5771366 article EN 2011-03-01

Recently, technologies such as face detection, facial landmark localisation and recognition verification have matured enough to provide effective efficient solutions for imagery captured under arbitrary conditions (referred "in-the-wild"). This is partially attributed the fact that comprehensive "in-the-wild" benchmarks been developed recognition/verification. A very important technology has not thoroughly evaluated yet deformable tracking "in-the-wild". Until now, performance mainly...

10.1007/s11263-017-0999-5 article EN cc-by International Journal of Computer Vision 2017-02-25

The human face is a rich source of information for the viewer and facial expressions are major component in judging person's affective state, intention personality. Facial an important part human-human interaction have potential to play equally human-computer interaction. This paper evaluates various active appearance model (AAM) fitting methods, including both original formulation as well several state-of-the-art task automatic expression recognition. AAM powerful statistical modelling...

10.1109/acii.2009.5349489 article EN 2009-09-01

We propose a face alignment framework that relies on the texture model generated by responses of discriminatively trained part-based filters. Unlike standard models built from pixel intensities or generic filters (e.g. Gabor), our has two important advantages. First, virtue discriminative training, invariance to external variations (like identity, pose, illumination and expression) is achieved. Second, we show (or patch-experts) are sparse can be modeled using very small number parameters....

10.1109/tpami.2014.2362142 article EN IEEE Transactions on Pattern Analysis and Machine Intelligence 2014-10-09

This paper is concerned with the problem of estimating motion a single camera from sequence images, an application scenario vehicle egomotion estimation. Egomotion estimation has been active area research for many years and various solutions to have proposed. Many methods rely on optical flow or local image features establish spatial relationship between two images. A new method presented which makes use Fourier-Mellin Transform registering images in video sequence, rotation translation can...

10.1109/ivs.2007.4290156 article EN IEEE Intelligent Vehicles Symposium 2007-06-01

Face recognition in real-world conditions requires the ability to deal with a number of conditions, such as variations pose, illumination and expression. In this paper, we focus on head pose use computationally efficient regression-based approach for synthesising face images different poses, which are used extend training set. data-driven approach, correspondences between facial landmark points frontal non-frontal views learnt offline from manually annotated data via Gaussian Process...

10.5244/c.23.31 article EN 2009-01-01

We present a novel approach to pose-invariant face recognition that handles continuous pose variations, is not database-specific, and achieves high accuracy without any manual intervention. Our method uses multidimensional Gaussian process regression learn nonlinear mapping function from the 2D shapes of faces at non-frontal corresponding frontal shapes. use this take an input image new arbitrary pose-normalize it, generating synthetic then used for recognition. fully automatic system...

10.5244/c.25.127 article EN 2011-01-01

We propose a 3D Constrained Local Model framework for deformable face alignment in depth image. Our exploits the intrinsic geometric information data by utilizing robust histogram-based features that are based on normal vectors. In addition, we demonstrate fusion of intensity and further improves facial landmark localization accuracy. The experiments conducted publicly available FRGC database. results show our CLM completely outperforms raw term fitting accuracy robustness, feature...

10.1109/icip.2014.7025285 article EN 2022 IEEE International Conference on Image Processing (ICIP) 2014-10-01

We present a robust real-time face tracking system based on the Constrained Local Models framework by adopting novel regression-based Discriminative Response Map Fitting (DRMF) method. By exploiting algorithm's potential parallelism, we hybrid CPU-GPU implementation capable of achieving performance at 30 to 45 FPS, ordinary consumer-grade computers. have made software publicly available for research purposes

10.1145/2557642.2579369 article EN 2014-03-19

The issue of transferring facial expressions from one person's face to another's has been an area interest for the movie industry and computer graphics community quite some time. In recent years, with proliferation online image video collections web applications, such as Google Street View, question preserving privacy through de-identification gained in vision community. this paper, we focus on problem real-time dynamic expression transfer using Active Appearance Model framework. We provide...

10.1109/icpr.2010.923 article EN 2010-08-01

The issue of transferring facial performance from one person's face to another's has been an area interest for the movie industry and computer graphics community quite some time. In recent years, deformable models, such as Active Appearance Model (AAM), have made it possible track synthesize faces in real Not surprisingly, model-based approaches transfer gained tremendous vision community. this paper, we focus on problem real-time using AAM framework. We propose a novel approach learning...

10.1109/tvcg.2011.157 article EN IEEE Transactions on Visualization and Computer Graphics 2011-09-28

Whilst deep learning techniques have achieved excellent emotion prediction, they still require large amounts of labelled training data, which are (a) onerous and tedious to compile, (b) prone errors biases. We propose Multi-Task Contrastive Learning for Affect Representation (MT-CLAR) few-shot affect inference. MT-CLAR combines multi-task with a Siamese network trained via contrastive infer from pair expressive facial images the (dis)similarity between expressions, difference in valence...

10.1145/3581783.3613784 article EN 2023-10-26

Statistically motivated approaches for the registration and tracking of non-rigid objects, such as active appearance model (AAM), have become very popular. A major drawback these is that they require manual annotation all training images which can be tedious error prone. In this paper, a MPEG-4 based approach automatic frontal face images, having any arbitrary facial expression, from single annotated image presented. This utilises animation system to generate virtual different expressions...

10.1109/icip.2009.5414123 article EN 2009-11-01

Statistical approaches for building non-rigid deformable models, such as the active appearance model (AAM), have enjoyed great popularity in recent years, but typically require tedious manual annotation of training images. In this paper, a learning based approach automatic visually objects from single annotated frontal image is presented and demonstrated on example automatically annotating face images that can be used AAMs fitting tracking. This employs idea initially correspondences between...

10.1109/cvpr.2009.5206766 article EN 2009 IEEE Conference on Computer Vision and Pattern Recognition 2009-06-01

Feature shaping refers to a family of methods that exhibit state-of-the-art performance for out-of-distribution (OOD) detection. These approaches manipulate the feature representation, typically from penultimate layer pre-trained deep learning model, so as better differentiate between in-distribution (ID) and OOD samples. However, existing feature-shaping usually employ rules manually designed specific model architectures datasets, which consequently limit their generalization ability. To...

10.48550/arxiv.2402.00865 preprint EN arXiv (Cornell University) 2024-02-01

Quantized networks use less computational and memory resources are suitable for deployment on edge devices. While quantization-aware training (QAT) is a well-studied approach to quantize the at low precision, most research focuses over-parameterized classification with limited studies popular device friendly single-shot object detection semantic segmentation methods like YOLO. Moreover, majority of QAT rely Straight Through Estimator (STE) approximation which suffers from an oscillation...

10.1109/wacv57701.2024.00244 article EN 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2024-01-03

We propose a learning-based approach to segment the seminal vesicles (SV) via random forest classifiers. The proposed discriminative relies on decision using high-dimensional multi-scale context-aware spatial, textual and descriptor-based features at both pixel super-pixel level. After affine transformation template space, relevant are extracted classifiers learned based masked region of from most similar atlases. Using these classifiers, an intermediate probabilistic segmentation is...

10.1117/12.2043893 article EN Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2014-03-21

There are various scenarios where finding the most similar expression is requirement rather than classifying one into discrete, pre-defined classes, for example, facial transfer and based automatic album generation. This paper proposes a novel method expression. Instead of regular L2 norm distance, we investigate use Structural SIMilarity (SSIM) metric similarity comparison as distance in nearest neighbour unsupervised algorithm. The feature vectors generated using Active Appearance Models...

10.1109/fg.2011.5771354 article EN 2011-03-01
Coming Soon ...