Clothing Instance Segmentation in the Wild: A Supervised Deep Learning Approach

Research Thesis
MSc in Data Science and Business Analytics, Bocconi University
Supervisor: Francesco Grossetti

Goal

This repo contains a pipeline to perform clothing instance segmentation on wild images: identifying with segmentation masks 30 fashion items in un-constrained real-world images, referred to as "wild" and retrieved from Instagram. Wild images are the opposite of "street" images that are out-of-the-studio good quality pictures, usually focused on one professional model, with minor occlusion and clear backgrounds.

To solve the task, a supervised approach has been implemented, with Mask R-CNN as segmentation model. The train and validation datasets have been created using DeepFashion2, a publicly-available already-labeled fashion dataset with wild and street photos. The test set has been created completely from scratch using random photos downloaded from Instagram. To improve the training’ speed and accuracy, transfer learning has been performed from COCO instance segmentation.

Requirements

Create a virtual environment and install the required libraries. Make sure your pip version is updated.

conda create -n venv python==3.9.0
conda activate venv
python -m pip install --upgrade pip
pip install -r requirements.txt 

Then install Detectron2, according to the steps from the official documentation, given the following requirements are satisfied.

1. Linux or macOS with Python ≥ 3.6
2. PyTorch ≥ 1.8 and torchvision that matches the PyTorch installation. Install them together at pytorch.org to make sure of this.
3. gcc & g++ ≥ 5.4

For example, on a Linux device with CUDA 11.3, run the following lines to install PyTorch and Detectron2:

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

Pipeline

Read this section to understand the purpose of each file in this repo.

1. Literature Review

To solve an instance segmentation problem in a supervised manner, two elements are required: images and annotations. Annotating images from scratch is really time consuming, thus the idea to build the train/validation set from one of the publicly-available, already-labeled fashion instance segmentation datasets: DeepFashion2, ModaNet, or Fashionpedia.

• summary_datasets.ipynb: contains code to download DeepFashion2 (DF2) and Fashionpedia, to analyze and visualize samples from DF2, ModaNet, Fashionpedia
• download_modanet.ipynb: contains code to download ModaNet
• converters.ipynb: contains code to convert json files from DF2 to COCO format (used in: summary_datasets.ipynb)

2. Dataset Creation

Given the goal to segment wild Instagram pictures, the train/validation set should contain both wild and street images. For this reason, images from DeepFashion2 (DF2) have been used to build three different versions of the train/validation set. The idea is to test how the same model, trained for each experiment on one version of the train set, performs on the same wild test set.

Version	Dataset	Source	# images
1	Train	Street images from DF2	2500
	Validation	Street images from DF2	500
2	Train	Wild images from DF2	5000
	Validation	Wild images from DF2	500
3	Train	Street + Wild images from DF2	7500
	Validation	Street + Wild images from DF2	1000
	Test	Wild images from Instagram	500

2.1 Data Collection and Cleaning

• to_annotate.ipynb: contains code to explain the collection and cleaning process for the train, validation, test sets
• color_detector.py: contains code to detect color vs. no-color images given a folder with images (used in: to_segment.ipynb)
• face_detector.py: contains code to detect face vs. no-face images given a folder with images (used in: to_segment.ipynb)

2.2 Data Annotation

Annotator: VGG Image Annotator (VIA)
Download Link

VGG Image Annotator (VIA) has four advantages over other annotation tools: it has no limits in terms of number of annotations; it can be downloaded and runs locally, without the need to upload any image anywhere; it gives users the possibility to draw annotations first and define labels in a second moment; it allows to define attributes at image level. The only problem of VIA is that the json file with annotations is not in the standard COCO format, but converting VIA annotations to COCO is easy.

Categories: 30 fashion categories have been considered. They have been retrieved from a fashion ontology, stored in tools/fashion_ontology.xlsx, that maps 211 fashion items to a supercategory, category and subcategory.

Label	ID	Label	ID	Label	ID
Hat	1	Pyjama	11	Leggings	21
Bag	2	Robe	12	Joggers	22
Boots	3	Dress	13	Jacket	23
Slippers	4	Jumpsuit	14	Blazer	24
Sneakers	5	Cover-ups	15	Waistcoat	25
Heels	6	Coat	16	Cardigan	26
Flat Shoes	7	Pants	17	Sweater	27
Swimsuit	8	Jeans	18	Sweatshirt	28
Bathrobe	9	Shorts	19	Shirt	29
Underwear	10	Skirt	20	Top	30

Annotation:
Train Set: re-annotate DF2 images according to 30 categories
Validation Set: re-annotate DF2 images according to 30 categories
Test Set: annotate IG images from scratch according to 30 categories
Annotations in COCO format for each set are available in data folder.

• to_annotate.ipynb: contains code to track annotation progress and manipulate json files from VIA
• pre_convert.ipynb: contains some code to manipulate json files in VIA format, before converting them to COCO format with the code in converters.ipynb
• copy_files.py: contains code to copy files from a folder in a new specified folder and rename them from 0 to N (used in: pre_convert.ipynb)
• converters.ipynb: contains code to convert json files from VIA to COCO format

2.3 Data Analysis

Given images and annotations, some metrics have been computed for each set to assess the quality of annotations (number of masks, average number of masks per image, average number of vertices per mask), the quality of images (average resolution, percentage of images with 1 person, with more than 1 person, without any person, percentage of selfies) and to understand the frequency and average dimension of fashion items (categories distribution, average mask area over image area).

• summary_mydataset.ipynb: contains code to compute all those metrics

3. Training and Inference

Mask R-CNN has been implemented as the instance segmentation model, using Detectron2, Facebook AI Research's python library for image segmentation and object detection.
Detectron2 Documentation

• detectron2.ipynb: contains code to understand how to install and use Detectron2 library, including how to train, evaluate and make inference with Mask R-CNN
• train_evaluate.py: contains code to train and evaluate Mask R-CNN, given a train, validation and test set (transfer learning is performed)
• experiments.ipynb: contains code to access the training and evaluation metrics, for each experiment