Neural-Vision Outfit Recommender

Ensemble CV + Flask API — cosine NN retrieval with demographic-aware search pruning.

A high-precision image recommendation engine that uses a multi-model ensemble for visual similarity, demographic markers, and occasion context.

Flask REST API, ONNX runtime, cosine NN retrieval with demographic pruning — sub-2s end-to-end.

Mission

Deliver outfit recommendations that go beyond pixel matching by combining ensemble computer vision, demographic-aware filtering, and fast cosine-similarity retrieval.

Innovation

Confidence-weighted consensus across InsightFace, DeepFace, and a custom FairFace-style ResNet34: when the custom head exceeds a high confidence threshold its vote wins; otherwise a majority vote sets demographic filters before nearest-neighbor search.

Technical deep dive

Architecture: ensemble of InsightFace, DeepFace, and a transfer-learned ResNet34 (FairFace-style) with dual classification heads for robust attribute extraction.
Consensus logic: prioritize ResNet34 when confidence > ~95%; else blend InsightFace 512D embeddings with DeepFace attribute signals via majority vote.
Retrieval: nearest neighbors with cosine similarity; search space pruned using predicted demographics and occasion metadata before similarity for relevance and speed.
Deployment: Flask REST API — raw image upload, GPU ONNX inference, Base64-encoded recommendation payloads for frontend integration.

Python
PyTorch
Flask
OpenCV
ONNX
InsightFace
Scikit-learn

Open Colab Let's talk