Publications

1. SSMTReID-Net: Multi-Target Unsupervised Domain Adaptation for Person Re-Identification (Spring ‘21)
   [Submitted for review to Pattern Recognition Letters]
   Advisor: Prof. Biplab Banerjee, CSRE, IITB
   Abstract: We tackle the problem of person re-identification (reID) and introduce a novel single–source multiple-target (SSMT) domain adaptation (DA) setup in this regard. Specifically, we consider a labeled source domain and multiple unlabeled target domains where a domain is defined in terms of the respective camera setup. Several recent reID works leverage off-the-shelf DA techniques to reduce any domain gap while deploying a reID system trained on a labeled source dataset into a unlabeled target domain under a single-source single-target fashion. However, such models have two inherent drawbacks if used for the single-source multi-target case: a) the performance of the source domain is hindered once the model is fine-tuned for a target domain, and b) posing the single–source multi-target setup as independent single-source single-target cases fails to utilize the complementary information present in the different target domains. To this end, we propose a novel architecture called SSMTReID-Net which considers the elastic weight consolidation (EWC) regularizer to ensure competitive performance of the source domain after adaptation and the notion of information bottleneck (IB) to highlight domain invariant target features while suppressing any task-irrelevant artifacts. The model is end-to-end trainable and results on different single-source multi-target combinations on the DukeMTMC-reID, Market-1501, and CUHK03 datasets confirm the superiority of SSMTReID-Net over the baselines.

2. FRIDA - Generative Feature Replay for Incremental Domain Adaptation (Autumn ‘20)
   [Accepted for publication at Computer Vision and Image Understanding]
   Advisor: Prof. Biplab Banerjee, CSRE, IITB
   Abstract: We tackle the novel problem of incremental unsupervised domain adaptation (IDA) in this paper. We assume that a labeled source domain and different unlabeled target domains are incrementally observed with the constraint that data corresponding to the current domain is only available at a time. The goal is to preserve the accuracies for all the past domains while generalizing well for the current domain. The IDA setup suffers due to the abrupt differences among the domains and the unavailability of past data including the source domain. Inspired by the notion of generative feature replay, we propose a novel framework called Feature Replay based Incremental Domain Adaptation (FRIDA) which leverages a new incremental generative adversarial network (GAN) called domain-generic auxiliary classification GAN (DGAC-GAN) for producing domain-specific feature representations seamlessly. For domain alignment, we propose a simple extension of the popular domain adversarial neural network (DANN) called DANN-IB which encourages discriminative domain-invariant and task-relevant feature learning. Experimental results on Office-Home, Office-CalTech, and DomainNet datasets confirm that FRIDA maintains superior stability-plasticity trade-off than the literature. [arXiv preprint]

3. Behavior of Keyword Spotting Networks Under Noisy Conditions (Summer ‘20)
   [Published at International Conference on Artificial Neural Networks 2021]
   Advisor: Prof. Oliver Bringmann, University of Tuebingen, Germany
   Abstract: Keyword spotting (KWS) is becoming a ubiquitous need with the advancement in artificial intelligence and smart devices. Recent work in this field have focused on several different architectures to achieve good results on datasets with low to moderate noise. However, the performance of these models deteriorates under high noise conditions as shown by our experiments. In our paper, we present an extensive comparison between state-of-the-art KWS networks under various noisy conditions. We also suggest adaptive batch normalization as a technique to improve the performance of the networks when the noise files are unknown during the training phase. The results of such high noise characterization enable future work in developing models that perform better in the aforementioned conditions. [arXiv preprint]