Research thesis · ICCIT 2026 target
DRISHTI-Bn
A parameter-efficient multimodal framework for Bengali disaster-response understanding.
PyTorchViT-B/16BanglaBERTLoRAMultimodal
Overview
DRISHTI-Bn explores flood damage classification and Bengali caption generation for a low-resource, disaster-response setting. The project combines a ViT-B/16 vision encoder with BanglaBERT and LoRA-based adaptation.
My contribution
- Designed the multimodal research direction and training workflow.
- Prepared Bengali annotations from CrisisMMD flood imagery using machine-translated labels.
- Investigated parameter-efficient adaptation with LoRA rather than full-model fine-tuning.
Approach and tools
- Visual representation: ViT-B/16.
- Bengali language representation: BanglaBERT.
- Adaptation strategy: LoRA with r=8 and alpha=16.
- Dataset context: 3,755 CrisisMMD flood images with Bengali annotations.
Evidence and status
- Architecture and ablation results will be added when the thesis experiments are finalized.
- The manuscript is in preparation for a possible ICCIT 2026 submission.
Limitations
Machine-translated Bengali annotations and the limited disaster-domain dataset require careful evaluation before real-world deployment.
Next steps
- Complete baseline and ablation comparisons.
- Report classification and caption-generation metrics.
- Document reproducibility, ethics, and dataset limitations.