Segmentation Task

Toy Project

• Topic: Human Segmentation using U-Net model with Kaggle image datasets
• Model: U-Net: Convolutional Networks for Biomedical Image Segmentation paper

Demo Video

0. Install Requirements

pip install -r requirements.txt

1. Download Dataset

• Sketch Dataset
• EDA
  • Images: Extracted from video frames
    • Number: 2615
    • Shape: (3, 960, 540)
  • Masks: Segmented images of dancing people
    • Number: 2615
    • Shape: (1, 960, 540)
  • Features:
    • No annotation file. Only masked image files.
    • Not split into train, validation, and test datasets.

2. Training And Validation

python train.py --batch_size 8 --epoch 35

3. Inference

Download Best.pt

• Best.pt

python infer.py

Results

Prediction

Training Loss Results

Best IoU Results

Structure

unet_tiktok
│
├── assets
├── date            # Download dataset
│   ├── images 
│   └── masks
├── networks
|   └── model.py
├── new_results     # Resize Predicted masks
├── results         # Predicted masks
├── saves
|    └── 0000
|    └──...
├── app.py          # Demo with Gradio
├── utills
├── infer.py
├── show.py
├── test_img.png
└── train.py