Few-Shot Fine-Grained Recognition with Context
Inject geography/language priors into a CLIP-style baseline for fine-grained categories.
Outcome. Added lightweight context conditioning to a CLIP-like baseline, yielding consistent gains in few-shot regimes.
Role: Undergraduate researcher · Dates: Oct 2023 – Jan 2024 · Stack: PyTorch, HuggingFace, Pandas
Highlights
- Engineered geo/lang priors and fusion heads; ablated conditioning strength and prompt variants.
- Built stratified few-shot splits with seeded runs for full reproducibility.
- Reported improvements on fine-grained sets without sacrificing inference speed.
