Back to Dashboard

Image Classification Analysis:
FGVC-Aircraft Benchmark

Author

Vũ Minh Quân

Dataset

FGVC-Aircraft

Peak Accuracy

88.9%

Project Overview

This project focuses on Fine-Grained Visual Classification (FGVC) using the aircraft dataset. We perform a comparative analysis between ResNet50 (CNN) and ViT-B/16 (Transformer). Our results indicate that ViT, combined with Layer-wise Learning Rate Decay (LLRD), achieves superior accuracy by effectively capturing global structural dependencies.

Download Full Report (PDF)

1. Exploratory Data Analysis (EDA)

Analyzing the initial dataset to understand the distribution and morphological characteristics of 100 aircraft variants.

Dataset Summary
Figure 1.1: Dataset Summary

Distribution of samples across Train, Validation, and Test sets.

Manufacturer Distribution
Figure 1.3: Manufacturer Distribution

Aircraft distribution categorized by major manufacturers.

Split Distribution
Figure 1.4: Split Balance

Maintaining data balance across training stages.

UMAP Visualization
Figure 1.5: UMAP Feature Projection

Visualizing latent space features using UMAP dimensionality reduction.

Data Samples
Figure 1.6: Dataset Samples
Edge Detection
Figure 1.7: Structural Edge Analysis
Color Histograms
Figure 1.8: Color Distribution

2. Training Dynamics

Monitoring the convergence and performance comparison across different training strategies.

Accuracy Curves
Figure 1.16: Accuracy Convergence
Loss Curves
Figure 1.17: Loss Optimization
Performance Steps
Figure 1.16.5: Accuracy Progression
Metric Summary
Figure 1.18: Metric Distribution
Final Summary
Figure 1.19: Final Test Accuracy Summary

3. Visual Interpretability

Visualizing model attention using Grad-CAM and Attention Rollout to validate decision logic.

Grad-CAM
Figure 1.13: Grad-CAM (ResNet50)

Focus on local parts such as engines and wingtips.

Attention Rollout
Figure 1.14: Attention Rollout (ViT-B/16)

Global silhouette awareness and structural coherence.

4. Detailed Evaluation

In-depth error analysis and performance metrics across different variants and strategies.

Confusion Matrix ResNet
Figure 1.9: Confusion Matrix (ResNet50)
Confusion Matrix ViT
Figure 1.10: Confusion Matrix (ViT-B/16)
Error Analysis
Figure 1.11: Error Prediction Analysis
Class-wise Accuracy
Figure 1.12: Class-wise Accuracy Breakdown
Linear Probing
Table 1.2: Linear Probing Benchmarks
Full Fine-tuning
Table 1.3: Full Fine-tuning Metrics
LLRD Strategy
Table 1.5: LLRD Strategy Comparison