🌳 TreeSpicesClassification – Multimedia Data Processing Project

3D → 2D → Feature Extraction → Classical ML → Deep Learning → 3D Learning

📌 Project Overview

TreeSpicesClassification is a Multimedia Data Processing project that focuses on the classification of 3D tree leaf/branch models across several species using three major methodological families:

• Indirect Methods - Convert 3D to 2D, extract classical features
• Quasi-Direct Methods - Multi-view 2D deep learning
• Direct 3D Deep Learning Methods - Direct point cloud processing

The goal is to compare how different feature extraction, representation, and learning techniques perform on the same 3D dataset.

Key Details:

• The dataset contains 3D point clouds of 7 species of trees
• Processed models with 2D/3D feature extraction
• Trained classical ML and deep learning models
• Comprehensive evaluation and visualization of results

---

🏗️ Project Architecture

TreeClassification/
│
├── Utils/
│   ├── data_loader.py              # Load & preprocess point cloud data
│   ├── preprocessing.py            # Data cleaning & normalization
│   ├── projections.py              # 3D → 2D projection converter
│   ├── feature_extraction.py       # Extract LBP, PFH, FPFH, CNN features
│   └── helpers.py                  # Utilities & visualization
│
├── Datasets/
│   ├── Train/                      # 80% training split
│   └── Test/                       # 20% testing split
│
├── Notebooks/
│   ├── LBP_Feature_Extraction.ipynb
│   ├── SVM_Classification.ipynb
│   ├── CNN_From_Scratch.ipynb
│   ├── ResNet_FineTuning.ipynb
│   ├── ResNet_FeatureExtraction_SVM.ipynb
│   ├── ResNet_TransferLearning.ipynb
│   ├── Fusion_CNN.ipynb
│   ├── QuasiDirect_PFH_SVM.ipynb
│   ├── Direct_PointNet.ipynb
│   └── Direct_DGCNN.ipynb
│
├── Outputs/
│   ├── MultiView_Data/
│   │   └── (5 projections per 3D model)
│   │
│   ├── Features/
│   │   ├── 2D_LBP_Features.npy
│   │   ├── 2D_FPFH_Features.npy
│   │   └── Features.csv
│   │
│   ├── Features3D/
│   │   ├── PFH_3D.npy
│   │   ├── FPFH_3D.npy
│   │   └── Descriptors/
│   │
│   └── Models/                     # (for future model saving)
│
└── README.md

---

🌱 Dataset Description

The dataset contains 3D point cloud files for 7 different tree species.

Data Distribution

All samples were preprocessed and split into:

• Train: 80%
• Test: 20%

Dataset Location

Datasets/Train/     # Training data (80% split)
Datasets/Test/      # Testing data (20% split)

Additional intermediate datasets (2D projections, extracted features) are saved under Outputs/.

---

🧪 Methods & Experiments

Below are the three large families of methods implemented in the notebooks.

🔹 1. Indirect Methods (2D Feature Extraction + Classical ML)

These methods convert 3D point clouds into 2D images, then extract classical features.

Steps:

1. Convert 3D point cloud → 2D projection images
2. Extract 2D features:
   • LBP (Local Binary Patterns)
   • HOG (if used)
3. Flatten & normalize features
4. Train classical ML models:
   • SVM (primary model used)
   • KNN / Random Forest (optional)

Related Notebooks:

• LBP_Feature_Extraction.ipynb
• SVM_Classification.ipynb

Outputs Saved: Located in Outputs/Features/:

• 2D_LBP_Features.npy
• Feature matrices (.npy / .csv)

---

🔹 2. Quasi-Direct Methods (Multi-View 2D Deep Learning)

Quasi-direct methods preserve partial 3D information through multi-view projection.

Steps:

1. Generate 5 projections per 3D model (top, bottom, side, oblique, etc.)
2. Train 2D CNN models:
   • CNN from Scratch
   • ResNet (Transfer Learning)
   • ResNet (Fine-Tuning)
3. Extract ResNet deep features and classify with SVM
4. (Optional) Fusion of multi-view CNN predictions

Related Notebooks:

• CNN_From_Scratch.ipynb
• ResNet_FineTuning.ipynb
• ResNet_FeatureExtraction_SVM.ipynb
• ResNet_TransferLearning.ipynb
• Fusion_CNN.ipynb

Outputs Saved: Located in Outputs/MultiView_Data/:

• 5 projected images per 3D object

Located in Outputs/Features/:

• Feature matrices extracted using CNN / ResNet

---

🔹 3. Direct Methods (3D Deep Learning)

These methods operate directly on 3D point cloud data without converting them to 2D.

State-of-the-Art Architectures:

• PointNet - Pioneering direct point cloud processing
• DGCNN (Dynamic Graph CNN) - Graph-based point cloud learning

Steps:

1. Load point cloud file
2. Normalize and sample points
3. Train PointNet/DGCNN
4. Evaluate on 20% test set

Related Notebooks:

• Direct_PointNet.ipynb
• Direct_DGCNN.ipynb

Outputs Saved: Located in Outputs/Features3D/:

• PFH / FPFH 3D descriptors
• 3D deep features
• Normalized point cloud samples

---

🧰 Utilities (Utils Folder)

The Utils/ directory contains reusable functions:

File	Purpose
data_loader.py	Load point cloud data, normalize, split
preprocessing.py	Denoising, normalization, format conversion
projections.py	Convert 3D → 2D (multi-view generator)
feature_extraction.py	Extract LBP, PFH, FPFH, CNN features
helpers.py	Plotting, metrics, model utilities

---

📊 Outputs Summary

✔ Outputs/MultiView_Data/

Contains all 2D projections of the 3D dataset

• 5 images per 3D object
• Used by CNN/ResNet models

✔ Outputs/Features/

LBP, FPFH, CNN extracted features (.npy, .csv formats)

• Classical ML feature matrices
• Deep learning embeddings

✔ Outputs/Features3D/

3D features for direct learning methods

• PointNet embeddings
• PFH / FPFH descriptors
• Normalized point clouds

✔ Outputs/Models/

(Models not saved — folder available for future expansion)

---

🧠 Evaluation

Evaluation Metrics

Evaluation metrics applied across all methods:

• Accuracy - Overall correctness
• Precision / Recall / F1-score - Per-class performance
• Confusion Matrix - Error analysis
• Training Time - Computational efficiency
• Robustness - Noise, rotation, point density variations

Comparative Analysis

Method	Type	Description	Scope
Indirect	2D + Classical ML	LBP + SVM	Fast, simple
Quasi-Direct	Multi-view CNN	ResNet, fine-tuning, fusion	Strong performance
Direct	3D Deep Learning	PointNet, DGCNN	Highest geometric fidelity

---

🚀 How to Use

1️⃣ Prepare Environment

pip install -r requirements.txt

2️⃣ Explore Notebooks

Open any notebook in the Notebooks/ folder to run experiments:

• Choose based on the method you want to explore
• Each notebook is self-contained with explanations

3️⃣ Generate 2D Views (Optional)

python Utils/projections.py

This will generate 5 projections for each 3D model and save to Outputs/MultiView_Data/.

4️⃣ Extract Features

python Utils/feature_extraction.py

This will extract LBP, FPFH, and other features, saving results to Outputs/Features/.

5️⃣ Train Models

Choose the notebook depending on your method:

• Classical ML: SVM_Classification.ipynb
• 2D CNN: CNN_From_Scratch.ipynb
• ResNet Transfer Learning: ResNet_TransferLearning.ipynb
• 3D Direct - PointNet: Direct_PointNet.ipynb
• 3D Direct - DGCNN: Direct_DGCNN.ipynb

---

Supervision: Said Ohamouddou

Repository: https://github.com/nvcy/TreeClassification

---

🏁 Conclusion

TreeSpicesClassification demonstrates how data representation dramatically influences classification performance.

By comparing Indirect, Quasi-Direct, and Direct approaches, this project provides a complete view of the challenges and trade-offs in processing 3D multimedia data:

• 📊 Indirect methods offer simplicity and speed
• 🎨 Quasi-direct methods balance 3D information with 2D CNN advantages
• 🔷 Direct 3D methods leverage geometric fidelity for optimal performance

This systematic exploration enables researchers and practitioners to make informed decisions when choosing methodologies for 3D object classification tasks.

---

📧 Contact

For questions or inquiries, please reach out via the GitHub repository.

---

Last Updated: November 26, 2025