Visual Taxonomy - Kaggle Competition sponsored by Meesho

🎯 Competition Overview

• Goal: Predict product attributes from fashion images
• Challenge: Accurately classify attributes like color, pattern, and sleeve length using only product images
• Final Ranking: 27th on the leaderboard

📊 Dataset Details

• Categories: 5 fashion categories
• Data Files:
  • category_attributes.parquet: Contains attribute names for each category
  • train.csv: Training data with product IDs, categories, and attributes
  • test.csv: Test data with product IDs and categories
  • sample_submission.csv: Submission template
• Image Data: Images stored as {product_id}.jpg in the images/ folder

🛠️ Technical Stack

• Deep Learning Models: ResNet101V2, EfficientNetB0, ConvNeXt Base
• ML Framework: CatBoost with category-specific models
• Preprocessing: SMOTE, SMOTETomek, Class Weight Balancing
• Libraries: PyTorch, TensorFlow, scikit-learn, TIMM

Feature Extraction

• Multi-model approach using:
  • ResNet101V2
  • EfficientNetB0
  • ConvNeXt Base

Preprocessing Techniques

• Visual similarity-based imputation
• Advanced feature scaling
• Handling class imbalance with SMOTE

Model Architecture

• Category-specific models
• CatBoost classification
• Adaptive training strategies

🧠 Project Structure

Meesho_Data_Challenge/
│
├── notebooks/
│   ├── data-exploration.ipynb
│   ├── model-training-demo.ipynb
├── src/
│   ├── __init__.py
│   ├── feature_extraction.py
│   ├── data_preprocessing.py
│   ├── models/
│   │   ├── __init__.py
│   │   ├── base_model.py
│   │   ├── mens_tshirts_model.py
│   │   ├── sarees_model.py
│   │   ├── kurtis_model.py
│   │   ├── womens_tshirts_model.py
│   │   └── womens_tops_model.py
│   └── pipeline.py
│
├── download_kaggle_data.py
├── main.py
├── requirements.txt
└── README.md