COMP9517 Group Project — Fine-Grained Species Classification

Overview

This repository contains the implementation, experiments, and report materials for the COMP9517 (Computer Vision) group project at UNSW Sydney, Term 2 2026. The project addresses fine-grained visual classification of living organisms — plants, animals, and fungi — using a subset of the iNaturalist-2021 dataset.

Fine-grained species recognition is a core challenge in biodiversity monitoring and citizen science: many species are visually near-identical, differing only in subtle morphological cues, while images of the same species vary widely in pose, lighting, background, and scale. This project develops and rigorously compares multiple computer vision pipelines — from classical handcrafted-feature methods to deep convolutional architectures — to tackle this problem at scale.

Objectives

• Build and evaluate at least two structurally different classification pipelines: a traditional handcrafted-feature approach (e.g., SIFT / HOG / LBP with a classical classifier) and a deep learning approach (CNN trained from scratch and via transfer learning).
• Benchmark methods using top-1 / top-5 accuracy, macro-averaged precision/recall/F1, and confusion-matrix-based error analysis.
• Investigate model interpretability using Grad-CAM to verify that predictions are driven by the organism itself rather than background context.
• Conduct in-depth ablation and robustness studies to understand failure modes and generalisation behaviour under real-world conditions (e.g., noise, blur, compression).

Dataset

• Source: iNaturalist-2021 "Mini" split (train_mini, val)
• Subset used: ≥ 500 randomly sampled species, with 40 training / 10 validation images per class, and 10 held-out test images per class (drawn from the official validation set)
• Data integrity: Training, validation, and test splits are kept strictly separate throughout all experiments to prevent data leakage.

Exact class lists, sampling seeds, and per-class image counts used in this project are documented in data/README.md for full reproducibility.

Methods

Category	Approach	Status
Traditional	Bag-of-Visual-Words (SIFT) + SVM	DONE
Traditional	HOG / LBP + Random Forest	DONE
Deep Learning	AlexNet trained from scratch and pretrained weights (transfer learning)	DONE
Deep Learning	ResNet 18 and ResNet50 trained from scratch and pretrained weights (transfer learning)	DONE
Advanced Method Development	Grad-CAM analysis,robustness to image degradation at test time,effect of the number of classes	DONE

Repository Structure

COMP9517GROUP_SUBJECT/
├── data/                  # Dataset scripts, class splits, and metadata (raw images not included)
├── notebooks/             # Exploratory analysis and prototyping notebooks
├── src/
│   ├── traditional/       # Handcrafted-feature pipelines (SIFT, HOG, LBP, SVM, RF)
│   ├── AlexNet/           # AlexNet architectures, training, and fine-tuning scripts
│   ├── ResNet/            # ResNet architectures, training, and fine-tuning scripts
│   ├── evaluation/        # Metrics, confusion matrix, macro-F1 computation
│   └── explainability/    # Grad-CAM and visualisation utilities
├── configs/                # Experiment configuration files
├── results/                 # Logs, metrics, figures (checkpoints and images excluded)
├── report/                # LaTeX (CVPR format) source for the written report
└── README.md

Getting Started

Requirements

python >= 3.10
pip install -r requirements.txt

Setup

git clone <repo-url>
cd COMP9517GROUP_SUBJECT
pip install -r requirements.txt

Downloading the Data

Refer to data/README.md for instructions on downloading the iNaturalist-2021 mini subset and generating the project's fixed 500-species split.

Running Experiments

#TODO: change it to running python script destination

# Traditional pipeline
python src/traditional/train.py --config configs/bovw_svm.yaml

# Deep learning pipeline
python src/AlexNet/train.py --config configs/resnet_pretrained.yaml

Results

Quantitative results, training curves, and confusion matrices will be added to results/ as experiments are completed, with a summary reported in the final written report and video presentation.

Team

Name	Contribution Area
Anrui Geng	Traditional computer vision pipeline using HOG, LBP, and SIFT-BoVW with SGD-SVM; ablation studies, result visualisation, and error analysis
Haochen Han	Traditional computer vision pipeline using HOG, LBP, and SIFT-BoVW with SGD-SVM; ablation studies, result visualisation, and error analysis
Sergio Insuasti	AlexNet trained from scratch and pretrained weights (transfer learning)
Zhirong Mai	ResNet trained from scratch and pretrained weights (transfer learning)
Zhiyang Liu	Grad-CAM explainability, robustness evaluation , test effect of the number of classes of traditional methods, report (LaTeX/CVPR) integration

References

• Van Horn et al., Benchmarking Representation Learning for Natural World Image Collections, CVPR 2021.
• Selvaraju et al., Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization, ICCV 2017.
• Hendrycks & Dietterich, Benchmarking Neural Network Robustness to Common Corruptions and Perturbations, ICLR 2019.

License

This repository is submitted as coursework for COMP9517 (UNSW Sydney) and is intended for academic evaluation purposes only.