A multipage demo inference application of different deep learning models built with Streamlit library. The app is built using Python 3.9.13. Here's the app
Following model/project demos are added here. It will be updated after each project.
- Cloth Segmentation Model
- Cloth Segmentation On Human Body Model
- Human Parsing Model
Here are some brief details of each model/project:
This is a binary semantic segmentation model trained on a dataset of single clothing article images. To specify, I used the cloth images and masks from the VITON PLUS dataset.
I used the MANet-EfficientnetB2 architecture for this and trained it for 40 epochs on the dataset. The model metrics are:
'test_DICE': 0.992,
'test_IOU': 0.984,
'test_TREVSKY': 0.989,
'test_loss': 0.0265
The model is trained with pytorch and converted to onnx (with optimization) to speed up inference.
This is a multi class semantic segmentation model to segment cloths from upper and lower part of the body . For this model, I have used the DeepFashion2 dataset. I have modified this dataset to have only upper and lower body cloths rather than the original dataset labels. This version of the dataset has 126750 training,3250 validation and 10000 test person images and corresponding pixel level annotations. The dataset has 3 different labels. The class labels are background,upper_body_cloth and lower_body_cloth.
Current model is based on UNET++-EfficientnetB3 architecture and trained for 18 epochs with the Categorical cross-entropy loss function. The model is trained with pytorch and converted to onnx (with optimization) to speed up inference. the current model metrics are:
'test_IOU': 0.746,
'test_DICE': 0.841,
'test_loss': 0.1441
This is a multi class semantic segmentation model to solve the human parsing task. For this model, I have used the
Look Into Person (LIP) dataset. This version of the dataset has 25000 training,
2500 validation and 2500 test multi person images and corresponding pixel level annotations. The dataset has 20 different human parts annotated. The class
labels are given below. You can read more about it from here.
Current model is based on UNET++-EfficientnetB3 architecture and trained for 24 epochs. The model is trained with pytorch and converted to onnx (with optimization) to speed up inference.
Note: This project is still ongoing so the metrics aren't provided as of now.