Fashion Item Detection
Key Results
- › 92% average precision across 19 clothing categories
- › Led a cross-functional team of 3 engineers
- › RetinaNet detection pipeline, served via FastAPI
- › Archived — the source data didn't transfer to the target market
The Problem
A fashion retail platform in the group wanted to recommend complete outfits from whatever item a customer was browsing. Rule-based recommendations had low engagement and no notion of visual style compatibility, so the first thing needed was a model that could actually see what was in a product image.
The approach
I led a team of three engineers building the detection layer: a RetinaNet pipeline that identifies and classifies garments in product images across 19 clothing categories, with OpenCV for preprocessing and augmentation, containerised with Docker and served over FastAPI.
What happened
The model reached 92% average precision across the 19 categories — a solid result for the detection problem, and worth being precise about: that number is detection precision, not a recommendation quality metric, and not a business outcome.
The project was ultimately archived. It was trained largely on European fashion data, and that didn’t transfer well to the Vietnamese market it was meant to serve — the garment distributions and styles were different enough to undermine it. No click-rate or cross-sell lift was ever measured, because it didn’t reach customers.
The useful residue is the team-lead experience and a clear-eyed lesson in dataset transfer: a model can be accurate on its test set and still be the wrong model for the market it lands in.