Kumar Shubham

Kumar Shubham

Auto-generated profile, based on publicly available information

Kumar Shubham

International Institute of Information Technology Bangalore, IN

Research Focus & Metrics

Representative Publications

View All Publications
7 November 2022
Verified
Conference Paper
Open access
JTJinal Hitesh Thakkar
Jinal Hitesh Thakkar
PRPooja S B. Rao
Pooja S B. Rao
KSKumar Shubham
Kumar Shubham
VJVaibhav Jain
Vaibhav Jain
DJDinesh Babu Jayagopi
Dinesh Babu Jayagopi
Recent technological advancements have boosted the usage of virtual interviewing platforms where the candidates interact with a virtual interviewing agent or an avatar that has human-like behavior instead of face-to-face interviews. As a result, it is essential to understand how candidates perceive these virtual interviewing avatars and whether adding features to boost the system’s interaction makes a difference. In this work, we present the results of two studies in which a virtual interviewing avatar with verbal and non-verbal interaction capabilities was used to conduct employment interviews. We add two interactive capabilities to the avatar, namely the non-verbal gestures and the verbal follow-up questioning and compare it with a simple interviewing avatar. We analyze the differences in perception with self-rated measures and behaviour with automatically extracted audiovisual behavioural cues. The results show that the candidates speak for a longer time, feel less stressed and have a better chance to perform with verbally and non-verbally expressive virtual interviewing agents.

8 December 2022
Verified
Conference Paper
Open access
SMSwasti Shreya Mishra
Swasti Shreya Mishra
KSKumar Shubham
Kumar Shubham
DJDinesh Babu Jayagopi
Dinesh Babu Jayagopi
The use of generated behavioral videos, particularly in a controlled fashion, is an emerging research direction in multi-modal interaction research. Apart from speaking-behavior, listening-behavior understanding is an important problem. Listening behaviors play a crucial role in many important dyadic conversations in day-to-day life. The interviewer uses cues from the listener’s behavior to judge a given candidate. A recent study has demonstrated the use of one-shot deep-fake generation models like the first-order motion model (FOMM) to transfer the behavior of an actor onto any single facial image for listening behavior understanding. Such an experiment requires the creation of multiple motion-transfer videos for different participants. FOMM, compared to other DeepFake models, provides a unique option of transferring any actor’s motion to a target source image without training multiple models for every participant. In addition, the unsupervised keypoint learning process allows FOMM to generalize to different kinds of motions. However, to generate high-quality output, it makes a first-order approximation of any motion. This limits its application, especially for transformations consisting of multiple rigid bodies and non-rigid body-based motions, especially for facial movements. As a result, these models often generate artifacts around the lip and eye regions. To circumvent this issue, in our work we have proposed a hybrid model which can approximate a given motion using both first-order and zero-order motion. Compared to the FOMM, our approach gives the network the flexibility to decide on the type of approximation it wants to make for a given keypoint. Results show that our model not only prevents distortion for non-rigid body motion but also generates better quality output when compared with FOMM.

8 December 2022
Verified
Conference Paper
Open access
RBRavi Kiran Reddy Kotlo Budde
Ravi Kiran Reddy Kotlo Budde
KSKumar Shubham
Kumar Shubham
GVGopalakrishnan Venkatesh
Gopalakrishnan Venkatesh
SGSriram Gandikota
Sriram Gandikota
SKSarthak Khoche
Sarthak Khoche
GSGopalakrishnan Srinivasaraghavan
Gopalakrishnan Srinivasaraghavan
Semantic face editing of real-world facial images is an important application of generative models. Recently, several works have explored possible techniques to generate such modifications by utilizing the latent structure of pre-trained GAN models. However, such approaches frequently necessitate the training of an encoder network, which is a time-consuming and resource-intensive procedure. StyleALAE [17], a latent-space based autoencoder that can generate photorealistic images of high quality, could be a viable alternative to a GAN-based architecture. Unfortunately, the reconstructed image in StyleALAE does not preserve the identity of the input facial image. This limits the application of StyleALAE for semantic face editing of images. In our work, we propose a one-shot learning approach to address this problem in StyleALAE. Our work ensures that the identity of the reconstructed image in StyleALAE is the same as the given input image. We also present ways to efficiently use the latent space of the pre-trained model to create semantic modification over the reconstructed image within the proposed one-shot learning approach. Results shows that our approach can generate semantic modifications on any real-world facial image while preserving identity.