PASCAL_VOC_ObjectClassification

Overview

This repository contains a MATLAB-based implementation of an object recognition pipeline developed for the PASCAL Visual Object Classes (VOC) Challenge. The project was carried out within the course Imaging Analysis and Processing (AIPI).

The task addressed is multi-class object presence classification on the PASCAL VOC 2006 dataset, which contains images with complex scenes, occlusions, and high intra-class variability.

The system follows the official PASCAL VOC protocol and evaluates image-level object classification across 10 categories: bicycle, bus, car, motorbike, cat, cow, dog, horse, sheep, and person.

---

Workflow Overview

---

Repository Structure

The repository is organized as follows:

• src/
  MATLAB source code implementing data loading, feature extraction, classifier training, and evaluation.

• data/
  Expected location of the PASCAL VOC Development Kit (not included).

• results/
  Evaluation results organized by CNN backbone, classifier, and feature layer. A dedicated results/README.md documents the results structure in detail.

• README.md
  Project-level documentation (this file).

---

Classification Strategy

The project adopts a two-stage hybrid strategy instead of end-to-end deep learning:

1. Deep Learning for Feature Extraction (Transfer Learning)

• Pretrained CNNs are used to extract feature embeddings from intermediate layers.
• Networks are pretrained on ImageNet.
• No fine-tuning or backpropagation is performed.
• CNNs are treated as fixed feature extractors.

Architectures evaluated include:

• AlexNet
• ResNet101
• EfficientNet
• MobileNet

2. Classical Machine Learning for Classification

• Extracted deep features are used to train supervised classifiers:
  • Support Vector Machine (SVM)
  • K-Nearest Neighbors (KNN)
• Classification follows a one-vs-rest scheme, with one binary classifier per object class.

This design enables systematic comparison of:

• CNN backbones
• Feature extraction layers
• Classifier behavior

CNN Backbones

Deep Learning Architectures Overview

Architecture	Total Parameters	Depth (Number of Layers)	Generalization Capacity	Computational Requirements
AlexNet	Medium	8	Low–Medium	Low
ResNet101	High	101	High	High
EfficientNet	Moderate	Variable (version-dependent)	Very High	Optimized
MobileNet	Low	Medium	Medium	Very Low

---

Practical Issues, Difficulties, and Solutions

Representative Feature Selection

Problem: Classical descriptors (SIFT, HOG) showed weak performance on VOC images due to occlusions and high variability.
Solution: Replaced handcrafted descriptors with pretrained CNN embeddings, which provided more robust and discriminative features.

Handling “Difficult” Objects

Problem: VOC annotations include samples marked as difficult, which distort training and evaluation.
Solution: Difficult samples were excluded from both training and validation, following VOC Development Kit recommendations.

File Structure and Data Access

Problem: VOC evaluation scripts require a strict directory layout.
Solution: The official VOC folder structure was enforced and all required paths were configured via VOCinit.m.

Computational Efficiency

Problem: CNN feature extraction is computationally expensive.
Solution: Extracted features were cached to disk, training was limited to train/val splits, and minibatch inference was used when possible.

One-vs-Rest Classifiers

Problem: VOC evaluation requires consistent binary classifiers per class.
Solution: Independent classifiers were trained per class using the official <class>_train.txt and <class>_val.txt splits.

---

Evaluation Protocol and Metrics

• Dataset: PASCAL VOC 2006
• Task: Image-level object classification
• Metrics:
  • ROC (Receiver Operating Characteristic) curves
  • AUC (Area Under the Curve)

AUC is used as the primary scalar metric:

• 1.0 indicates perfect classification
• 0.5 corresponds to random performance

---

Results Summary

Best Configurations

• AlexNet: fc7 + SVM
• ResNet101: pool5 + SVM
• MobileNet-v2: global_average_pooling2d_1 + SVM
• EfficientNet-b0: global_average_pooling2d + SVM

These configurations achieved consistently high AUC values across most object classes.

Intermediate Configurations

• Final or near-final layers (e.g. fc8, logits, fc1000) with SVM
• Good performance, but reduced generalization due to ImageNet specificity

Worst Configurations

• KNN-based classifiers across all CNNs
• Significant performance drops, especially for challenging categories such as person and horse

###Table Summary

CNN Backbone	Best Configuration (Layer + Classifier)	Intermediate Configuration	Worst Configuration
AlexNet	fc7 + SVM	fc8 + SVM	fc8 + KNN
ResNet101	pool5 + SVM	fc1000 + SVM	pool5 + KNN
MobileNet-v2	global_avg_pooling2d_1 + SVM	logits + SVM	global_avg_pooling2d_1 + KNN
EfficientNet-b0	global_avg_pooling2d + SVM	dense / MatMul + SVM	global_avg_pooling2d + KNN

---

Future Work and Key Takeaways

Key Takeaways

• Pretrained CNNs provide strong transferable representations for VOC-like datasets.
• SVM consistently outperforms KNN when classifying high-dimensional deep features.
• Intermediate CNN layers are more transferable than final classification layers.

Future Work

• Full fine-tuning of pretrained CNNs with access to stronger GPU resources
• Evaluation of newer architectures (EfficientNet-v2, Vision Transformers)
• Advanced data augmentation strategies
• Model ensembles and self-supervised learning approaches

---

References

• Everingham et al., The Pascal Visual Object Classes (VOC) Challenge, IJCV, 2010.
• Simonyan & Zisserman, Very Deep Convolutional Networks for Large-Scale Image Recognition, 2015.
• He et al., Deep Residual Learning for Image Recognition, CVPR 2016.
• PASCAL VOC Challenge: http://host.robots.ox.ac.uk/pascal/VOC/
• VLFeat Toolbox: http://www.vlfeat.org/
• PRTools Toolbox: https://prtools.tudelft.nl/