Fine-Grained Bird Species Classification on CUB-200 with ResNet-50

📋 Project Overview

We tackle the fine-grained classification of 200 bird species in the CUB-200-2011 dataset by designing an efficient, single-model pipeline based on ResNet-50. By systematically integrating:

• End-to-end: 542.py handles data loading, training and evaluation
• Key techniques: strong augmentations, OneCycleLR scheduling, label smoothing, gradient clipping, CBAM attention
• Performance: achieves 79.2 % Top-1 with stable convergence and minimal overfitting

we achieve 79.2 % ± 0.1 % Top-1 accuracy, with validation accuracy stabilizing above 77 % by epoch 20 and only minimal overfitting.

📁 Repository Contents

├── README.md           ← this document
├── final\_report.pdf    ← full write-up (motivation, methodology, results, conclusion)
├── slide.pptx          ← project presentation slides
├── assets/
│   └── videos/
│       └── demo.mp4    ← short demo of the trained model in action
├── midpoint/           ← code & experiments up to midpoint report
└── final/              ← polished scripts, logs, and model weights for final evaluation

📄 How to Explore

1. Read the report
   Open final_report.pdf for detailed descriptions of:

   • Problem statement & refined objectives
   • Related work & open-source references
   • Dataset preparation & preprocessing
   • Model architecture & training regimen
   • Quantitative results, learning curves & error analysis
   • Conclusions & future work

2. View the slides
   Open slide.pptx for a concise summary of our approach and key findings.

3. Inspect code and experiments

   • midpoint/ branch: scripts and notebooks used in the project midpoint.
   • final/ branch: final training scripts and weights matching the report’s results.

   # Switch between branches:
   git checkout midpoint   # midpoint experiments
   git checkout final      # final implementation
   

📚 Key References

• Wah et al., The Caltech–UCSD Birds-200-2011 Dataset, 2011.
• He et al., Deep Residual Learning for Image Recognition, CVPR 2016.
• Woo et al., CBAM: Convolutional Block Attention Module, ECCV 2018.
• Zhang et al., mixup: Beyond Empirical Risk Minimization, ICLR 2018.
• Zhong et al., Random Erasing Data Augmentation, AAAI 2020.
• Smith & Topin, Super-Convergence, 2019.

(See final_report.pdf bibliography for full citation details.)

🤝 Acknowledgements

• PyTorch & torchvision for ResNet-50 implementation
• jongchan/attention-module for CBAM reference code
• Facebook Research for the original MixUp implementation

---

Authors: Shiheng Xu · Renjie Fan · Kunshu Yang Date: May 2025