Department of Computer Science & Engineering

&ldquo;IIT Jodhpur is committed to impart quality education to facilitate the creation of qualified, competent and responsible human resources to meet the emerging technological challenges of the world. It is working towards creating an equitable learning platform that is positive, accessible and effective in the face of the modern societal and scientific advancements.

Flexibility in curriculum, in all its academic programs, encompassing elements of teaching, research, creativity, innovation and entrepreneurship is the most unique feature of IIT Jodhpur. The Department of Science and Technology (DST), under the National Mission on Cyber Physical Systems, has identified IIT Jodhpur as a Technology &amp; Innovation Hub in the area of Augmented Reality and Virtual Reality. The institute has also been nominated as the coordinator for the Jodhpur City and Knowledge Cluster by the Office of Principal Scientific Adviser, Government of India.

The sprawling 852-acre IIT Jodhpur campus is a majestic tribute to the deep-rooted diverse cultural heritage of the region. The campus is an architectural masterpiece, which is slated to become an international exemplar of sustainability with net-zero energy, and advanced water and waste management strategies.&rdquo;

IIT Jodhpur

Object tracking relies on a recursive search technique around the previous target location, concurrently learning the target appearance in each frame. A failure in any frame causes a drift from its optimal target path. Thus, obtaining highly confident search regions is essential in each frame. Motivated by the strong localization property of segmented object masks, the proposed method introduces instance segmentation as an attention mechanism in object tracking framework. The core contribution of this paper is threefold: (i) a region proposal module (RPM) based on instance segmentation to focus on search proposals having a high probability of being the target, (ii) a target localization module (TLM) to localize the final target using a correlation filter and (iii) a domain adaptation technique in both RPM and TLM modules to incorporate target specific knowledge and strong discrimination ability. Extensive experimental evaluation on three benchmark datasets demonstrate a significant average gain of 2.47% in precision, 2.55% in AUC score and 2.15% in overlap score in comparison with recent competing trackers. © 2021 Elsevier B.V.

World University root node

Image and Vision Computing

DA-SACOT: Domain adaptive-segmentation guided attention for correlation based object tracking

Computer Vision is the field of machine learning that deals with computers gaining knowledge from digital images/videos and performing tasks that human vision is capable of doing. It is widely used in the field of robotics for designing guidance systems where objects in the robot's field of view are identified and located. This research work is an application-specific project enabling a half-humanoid to find the 6D pose and bounding boxes of its hand and other objects within its field of view. We add an edge prediction head to the NOCS (Normalised Object Coordinate Space) model, which predicts the edges of each object from the predicted instance maps. An additional edge-agreement-loss found from the predicted edges is added to the total loss. This increases the attention to the edges and improves the accuracy of prediction of the instance masks. This edge-attention aided model is initialized with pre-trained weights of CAMERA and REAL dataset using transfer learning. The backbone layers of the model are frozen and the head layers alone are trained using a synthetic dataset (HAND dataset) we created using a software called blender. The model gives promising results when tested with objects kept in varying lighting conditions and at different distances from the camera. The use of transfer learning in models as large as the NOCS model allows us to train the model for a new class by only training the top few layers with a significantly small dataset. © 2021 IEEE.

2021 International Conference on Wireless Communications, Signal Processing and Networking, WiSPNET 2021

Faster Training of Edge-attention Aided 6D Pose Estimation Model using Transfer Learning and Small Customized Dataset

Modern deep learning systems have achieved unparalleled success and several applications have significantly benefited due to these technological advancements. However, these systems have also shown vulnerabilities with strong implications on the fairness and trustability of such systems. Among these vulnerabilities, bias has been an Achilles' heel problem. Many applications such as face recognition and language translation have shown high levels of bias in the systems towards particular demographic sub-groups. Unbalanced representation of these sub-groups in the training data is one of the primary reasons of biased behavior. To address this important challenge, we propose a two-fold contribution: a bias estimation metric termed as Precise Subgroup Equivalence to jointly measure the bias in model prediction and the overall model performance. Secondly, we propose a novel bias mitigation algorithm which is inspired from adversarial perturbation and uses the PSE metric. The mitigation algorithm learns a single uniform perturbation termed as Subgroup Invariant Perturbation which is added to the input dataset to generate a transformed dataset. The transformed dataset, when given as input to the pre-trained model reduces the bias in model prediction. Multiple experiments performed on four publicly available face datasets showcase the effectiveness of the proposed algorithm for race and gender prediction.

Fulltext

FRONTIERS IN BIG DATA

Subgroup Invariant Perturbation for Unbiased Pre-Trained Model Prediction

Due to the large applicability of AI systems in various applications, fairness in model predictions is extremely important to ensure that the systems work equally well for everyone. Biased feature representations might often lead to unfair model predictions. To address the concern, in this research, a novel method, termed as Attention Aware Debiasing (AAD) method, is proposed to learn unbiased feature representations. The proposed method uses an attention mechanism to focus on the features important for the main task while suppressing the features related to the sensitive attributes. This minimizes the model's dependency on the sensitive attribute while performing the main task. Multiple experiments are performed on two publicly available datasets, MORPH and UTKFace, to showcase the effectiveness of the proposed AAD method for bias mitigation. The proposed AAD method enhances the overall model performance and reduces the disparity in model prediction across different subgroups.

Proceedings of the IEEE International Conference on Computer Vision

Attention Aware Debiasing for Unbiased Model Prediction

US Patent 10,944,767

Identifying artificial artifacts in input data to detect adversarial attacks

INTERNATIONAL JOURNAL OF SCIENTIFIC STUDY

Disorders among Dentists in Punjab: A Questionnaire Study

In the scene of large crowd gatherings and challenging visiting places such as rough hills and high glass buildings, acquisition of the images through normal cameras is difficult and next to impossible. In all such scenarios, the drone becomes a useful acquisition sensor to capture the detailed information of the scene and the objects present there. With the rapid development of consumer unmanned aerial vehicles (UAV) or drones, the utilization of these devices became extremely easy. The popular use-case of the drones can be seen for the surveillance to identify any possible threats in the large crowd gathering and recognize the different individuals present in the crowd. However, the images captured using the drones are generally taken from a significant distance to avoid any collision; hence these images generally suffer in quality such as low resolution, motion blur and other environmental factors. The impact of these artifacts has been seen in the face recognition performance using several machine learning algorithms on large-scale drone databases namely Drone SURF. In this research, we intend to tackle the above artifacts by looking at the problem from the perspective of super-resolution of low-quality images. We have studied state-of-the-art (SOTA) super-resolution algorithms and see whether current methods are capable of handling the challenges of drone images. Apart from that, we have also evaluated another SOTA deep network developed for object detection for human segmentation in drone images. The proposed research provides interesting findings highlighting the limitations of existing works from the perspective of handling drone images. We would like the readers to go through the paper to find out the current limitations and possible future directions in drone image surveillance.

2021 IEEE International Workshop on Information Forensics and Security (WIFS {\ldots}

Impact of Super-Resolution and Human Identification in Drone Surveillance

Thirty-Fifth {AAAI} Conference on Artificial Intelligence, {AAAI} 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, {IAAI} 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, {EAAI} 2021, Vir

{{NEAP-F:} Network Epoch Accuracy Prediction Framework (Student Abstract)}

Aaai.Org

On Learning Deep Models with Imbalanced Data Distribution

Face verification involves identifying whether two faces belong to the same person or not. It relies heavily upon face perception, processing and the decision making of an individual. This research studies cross-domain face verification, where one face image belongs to a controlled, well-illuminated environment, while the other is of a varying representation having differences in image type or quality. Specifically, two cross-domain face verification tasks are analyzed: controlled-low resolution and controlled-sketch face verification. functional Magnetic Resonance Imaging (fMRI) data has been collected for 23 participants of two ethnic groups while performing face verification. Statistical comparisons were performed with same-domain controlled face verification for both the tasks. Our findings reveal regions of Right Frontal Gyrus, Bilateral Insula and Right Middle Cingulate Cortex demonstrating higher activation for controlled-sketch face verification, as compared to controlled face verification. Similar analysis were performed for controlled-low resolution face verification, where regions responsible for higher visual load and difficult tasks result in higher activation. Further, stimuli ethnicity differences influence activations for low-resolution face verification but do not affect sketch face verification. Regions of Right Middle Occipital Gyrus and Right Fusiform Gyrus present higher activity, suggesting increased face processing effort for within ethnicity low resolution face verification. We believe the findings of this research will help enable further development in the field of brain-inspired facial recognition algorithms.