Fashion Computer Vision

Overview

Built for the Artificial Neural Networks course at UAntwerp, applying the theory and best practices from the course to a small computer vision system for a fashion platform: classify clothing, find visually similar items, and keep it light enough to run on limited hardware.

Part A: Picking a backbone

Compared a standard CNN against a depthwise separable CNN. The separable version won outright: more accurate, and about 3x fewer parameters and FLOPs.

Part B: Learning similarity

Repurposed that backbone and trained it with triplet loss to produce embeddings for retrieval instead of just classification. Result: a correct match lands in the top-1 result ~84% of the time, and in the top-5 ~93% of the time.

Part C: Shrinking it further

Tried pruning and knowledge distillation on the retrieval model. Pruning preserved accuracy but gave no real speedup without sparse-aware hardware. Distillation traded some top-1 accuracy for a genuine ~3x smaller, ~3x cheaper model that still finds a correct match in its top-5 ~90% of the time.