Research

Computer Vision, Autonomous Vehicle, Deep Learning

View on Github

Undergraduate thesis project under Dr. Md. Monirul Islam. We implemented two novel pooling methods in the Convolutional Neural Network for semantic segmentation. The first method, wavelet pooling, balances computational efficiency with the preservation of fine features and edges, thereby enhancing segmentation accuracy in complex scenarios. The second method, pixel binning, strategically reduces feature map resolution to expand the receptive field, gathering more contextual information while maintaining computational efficiency. Evaluation on the CAMVID benchmark dataset demonstrates the effectiveness of both approaches, showcasing their competitive performance in terms of accuracy and efficiency compared to existing state-of-the-art methods. These advancements signify promising strides in refining semantic segmentation for applications crucial to autonomous vehicles and computer vision systems.

Reinforcement Learning, Ubiquitous Computing, Spatial Data Management

View on Github

Working with Dr. Tanzima Hashem. This work introduces the Green Route Optimization problem, focusing on finding the route with the minimum air and noise pollution and the maximum urban green space within a travel distance constraint. Using reinforcement learning, our approach adapts to real-time environmental changes, balancing environmental factors (i.e., air and noise pollution and green urban space) with travel efficiency. The key idea behind the effectiveness of our reinforcement learning based solution is our novel reward function that simultaneously considers the environmental factors and travel distance, and significantly enhance the quality of green routes. Our extensive experiment with real datasets show that our green routes reduces the pollution level and increases the green space with a reasonable increase of the travel distance.

Reinforcement Learning, Ubiquitous Computing, Spatial Data Management

View on Github

Worked with Dr. Md. Eunus Ali. The goal was to create a LitBank-like dataset for Bangla language. I have annotated some popular Bangla novels for this purpose using the Brat standoff annotation format, applying coreference resolution to identify all the pronouns with their respective nouns and generate tokens for each noun - pronoun pair. This work resulted in 6000 noun-pronoun pair tokens from different chapters of those novels, each 1500 words long