cartantino/FashionizerBot

A telegram bot that recognizes clothes in images and shows the clothes more similar to those contained in them

FashionizerBot

A telegram bot that recognize dresses and retrieve similar dresses

Run bot

python start_fashionizer.py

Handle dependencies easily

Conda

Create your own conda environment to run FashionizerBOT on your workstation:

In project root:

conda env create -f environment.yml
conda activate VIPM

Examples

Pipeline example

Classification example

Retrieval example

Hands on fashionizerBot backbone

Bot

• Basic implementation of the Bot picked up from dros1986/python_bot.
• start_fashionizer.py does all the works:
• Initialize the Updater that handles the Bot functions and input messages
• Loading of the MaskRCNN Resnet18 and SVM only once when the Bot starts

Segmentation

• MaskRCNN (MatterPort implementation, matterport/Mask_RCNN.)
• Segmentation to identify ROI and classify them as ['clothes', 'no clothes']

Results report

Classification

• Resnet18 finetuning
• Resnet18 features + SVM (Kernel = 'rbf', C = 1000, gamma = 1*e-3)
• BoW Features + SVM

Results report

Image retrieval

• KDTree from neural features
• KDTree from BOW features

Results report

Actual configuration after tests

• Segmentation : MaskRCNN
• Classification : Resnet18 neural features + SVM
• Retrieval : Resnet18 neural features + KdTree

Contacts

For any questions or doubt do not hesitate to contact us :

• Costantino Carta (c.carta4@campus.unimib.it)
• Hamza Amrani (h.amrani@campus.unimib.it)