nitesh.saini

Application

Image Data Cleaning Pipeline

Two-phase dedup and multi-LLM label validation for crowdsourced survey imagery.

15,197
Images
29.6%
Dupes removed
pHashHSVMulti-LLM consensusLVLM

Phase 1 removes exact and near-duplicates via hashing, HSV fingerprints, cosine similarity, and embeddings with multi-LLM verification. Phase 2 validates labels with rural-context calibration. Applied to 15,197 Odisha Income Baseline images.

Image Data Cleaning Pipeline
Focused · click to expand