louisJLN/clothes-stance

Purpose

This very lightweight model recognizes clothes pictures stance.

This model classifies clothes picture based on their stance. It can recognize picture of the front, back, label, close up and more. It is not perfectly accurate on the recognized stance but is perfect to sort pictures consistently e.g. front first, back second, third closeup and label last.

It has been trained on a dataset made up of 50% commercial studio pictures (from Fashionpedia) and 50% second hand ( =amateur) pictures.

The base model is nvidia_efficientnet_b0 that has been retrained using pytorch. You can see the training parameters here.

As the base model is super small (5.2M parameters), it runs fast on CPU and with a very small memory footprint. It also runs blazing fast on GPU. CUDA is not required to run the model.

Installation

pip install -r requirements.txt

Usage

See use-example.ipynb

Dataset

The dataset has been curated by hand. For any enquiry about the training dataset, contact me.