Unsupervised Domain Adaptation and Semi-Supervised Learning

Description: Build an end-to-end domain adaptation pipeline for the in-house DL team. For any classification problem, synthetic data was generated used or generated through publicly available datasets and local rendering which was then adapted on the real data image classification. Also designed a novel algorthim for semi-supervised learning built on top of the present algorithms such as FixMatch to obtain the best results on Long Tailed Datasets and Open Set Classification. Around 6% accuracy gains were observed on private datasets(10% data and 20 imbalance factor) after training the model using the proposed algorithm.