Siddhartha Gairola

DAVE : Distribution-aware Attribution via ViT Gradient Decomposition

Adam Wróbel, Siddhartha Gairola, Jacek Tabor, Bernt Schiele, Bartosz Zieliński, Dawid Rymarczyk

Pre-print, Feb. 2026

arxiv / code / presentation / bibtex

DAVE is a distribution-aware attribution method for Vision Transformers that produces stable, high-resolution pixel-level explanations while reducing patch/grid artifacts common in ViT saliency maps. It decomposes ViT input gradients to suppress operator-variation noise and enforces local equivariance via averaging over small transforms and perturbations, improving localization and faithfulness across multiple ViT backbones.

How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations

Siddhartha Gairola, Moritz Boehle, Francesco Locatello and Bernt Schiele

International Conference on Learning Representations (ICLR), 2025

OpenReview / arxiv / code / bibtex

Post-hoc importance attribution methods are a popular tool for "explaining" Deep Neural Networks (DNNs) and are inherently based on the assumption that the explanations can be applied independently of how the models were trained. Contrarily, in this work we bring forward empirical evidence that challenges this very notion. Surprisingly, we discover a strong dependency on and demonstrate that the training details of a pre-trained model's classification layer (<10% of model parameters) play a crucial role, much more than the pre-training scheme itself. With this finding we also present simple yet effective adjustments to the classification layers, that can significantly enhance the quality of model explanations.

SimPropNet: Improved Similarity Propagation for Few Shot Segmentation

Siddhartha Gairola, Ayush Chopra, Mayur Hemani and Balalji K.

International Joint Conferences on Artificial Intelligence (IJCAI), 2020

IJCAI_Proceedings / pdf / bibtex

Improving similarity propagation to improve one-shot and few-shot image segmentation.

Unsupervised Image Style Embeddings for Retrieval and Recognition Tasks

Siddhartha Gairola, Rajvi Shah, P.J. Narayanan

IEEE Winter Conference on Applications of Computer Vision (WACV '20), 2020

code / project_page / paper / supp_material / bibtex

An unsupervised protocol for learning a neural embedding of visual style of images. The proposed protocol does not leverage categorical labels but a proxy measure for finding stylistically similar and dissimilar images.

Find Me a Sky : A Data-driven Method for Color-Consistent Sky Search and Replacement

Saumya Rawat*, Siddhartha Gairola*, Rajvi Shah, P.J. Narayanan

International Conference on Multimedia Modelling (MMM '18), 2018

project_page / pdf / bibtex

A data driven method for color-consistent sky search and replacement. This technology does not require the use of complex color transfer techniques.

*Both authors contributed equally towards this work.

SmartKC++: Improving Performance of Smartphone-Based Corneal Topographers

Vaibhav Ganatra, Siddhartha Gairola, Nipun Kwatra and Mohit Jain et al.

IEEE Winter Conference on Applications of Computer Vision (WACV '25), 2025

Open Access / code / bibtex

Improving the SmartKC image processing pipeline, making it more robust and accurate.

Towards Automating Retinoscopy for Refractive Error Diagnosis

Aditya Aggarwal, Siddhartha Gairola, Nipun Kwatra and Mohit Jain et al.

Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT), Volume 6, Issue 3, 2022

code / project_page / pdf

In this work, we automate retinoscopy by attaching a smartphone to a retinoscope and recording retinoscopic videos with the patient wearing a custom pair of paper frames. We develop a video processing pipeline that takes retinoscopic videos as input and estimates the net refractive error based on our proposed extension of the retinoscopy mathematical model. Our system alleviates the need for a lens kit and can be performed by an untrained examiner.

Keratoconus Classifier for Smartphone-based Corneal Topographer

Siddhartha Gairola, Nipun Kwatra and Mohit Jain et al.

IEEE Engineering in Medicine & Biology Society (EMBC), 2022

arXiv / pdf

In this work, we propose a dual-head convolutional neural network (CNN) for classifying keratoconus on the heatmaps generated by SmartKC. Since SmartKC is a new device and only had a small dataset (114 samples), we developed a 2-stage transfer learning strategy to satisfactorily train our network, achieving a sensitivity of 91.3% and a specificity of 94.2%.

Smartphone based Corneal Topographer

Siddhartha Gairola, Nipun Kwatra and Mohit Jain

Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT), Volume 5, Issue 4, 2021

code / project_page / pdf / demo_video

SmartKC is a low-cost smartphone-based corneal topographer. It provides a low-cost solution for the mass screening of keratoconus at scale.

RespireNet: A Deep Neural Network for Accurately Detecting Abnormal Lung Sounds in Limited Data Setting

Siddhartha Gairola, Francis Tom, Nipun Kwatra and Mohit Jain

IEEE Engineering in Medicine & Biology Society (EMBC), 2021

code / arXiv / pdf / bibtex

RespireNet is a simple CNN-based model, along with a suite of novel techniques—device specific fine-tuning, concatenation-based augmentation, blank region clipping, and smart padding—enabling one to efficiently use small-sized datasets. We perform extensive evaluation on the ICBHI dataset, and improve upon the state-of-the-art results for 4-class classification by 2.2%.

Identifying Clickbait: A Multi-Strategy Approach Using Neural Networks

Dhruv Khattar, Siddhartha Gairola, Vaibhav Kumar, Yash Kumar Lal, Vasudeva Varma

ACM SIGIR, 2018

code / arxiv / bibtex

Detecting clickbaits using Deep Learning.