Projects
Projects
Ongoing and completed research tracks spanning long-tailed continual learning, CLIP adaptation, spatial VLM diagnostics, and robust AD/ADAS perception.
AdaPrior - Long-Tailed Continual Learning
Research problem: Class imbalance and prior drift in class-incremental settings lead to brittle predictions.
Key idea: Bayesian-inspired adaptive prior correction to reduce long-tail bias while preserving adaptation ability over incremental updates.
ViewDiag / Consistent Yet Wrong - Spatial VLM Evaluation
Research problem: Spatial reasoning models can be confident yet incorrect due to weak evidence grounding.
Key idea: Build diagnostics and benchmarks for evidence sensitivity to reveal failure modes in spatial VLM reasoning.
Robust Free-Space Perception for AD/ADAS
Research problem: Free-space and road segmentation fail under severe weather, lighting shifts, and sensor/domain changes.
Key idea: Multi-camera robust perception with temporal consistency constraints and deployment-aware optimization for embedded platforms.
Project/demo links will be shared once public assets are ready.
Continual / Incremental Learning in Remote Sensing
Research problem: Remote sensing pipelines require continual updates without catastrophic forgetting.
Key idea: Curriculum-driven incremental learning and informative subset selection for stable long-term visual recognition.