Object Recognition Project (COMP61342/41342: Cognitive Robotics and Computer Vision)

Author: Kunwei Song

Student ID: 11537231

1. Project Overview

This project investigates object recognition by comparing lightweight convolutional neural networks (CNNs) with a traditional computer-vision pipeline based on Bag-of-Visual-Words and SVM. Experiments are conducted on the iCubWorld1.0 and CIFAR-10 datasets to evaluate end-to-end learning versus handcrafted feature approaches.

2. Dataset

We use two publicly available datasets:

• iCubWorld1.0: A robotics vision dataset with images of 10 object categories captured under various conditions.

  • Training Samples: 600 images (3 object instances × 200 images)
  • Test Subsets: 4 partitions (Background, Categorization, Demonstrator, Robot)

• CIFAR-10: A standard benchmark for image classification containing 60,000 32×32 color images in 10 classes.

  • Training Samples: 50,000
  • Test Samples: 10,000

These datasets provide complementary challenges: iCubWorld1.0 tests robustness to background and viewpoint changes, while CIFAR-10 measures performance on small-scale natural images.

3. Methods Implemented

We implement and evaluate two main object-recognition pipelines:

Method 1: Bag-of-Visual-Words + SVM

• Feature Extraction: SIFT descriptors on densely sampled keypoints.
• Vocabulary Construction: K-means clustering to form a 500-word codebook.
• Encoding: TF–IDF weighting with L2 normalization and stop-word removal variants.
• Classification: Linear SVM on histogram representations.

Method 2: Convolutional Neural Networks (CNNs)

• SimpleCNN (Flatten): A small CNN with fully connected layers after convolution.
• SimpleCNN (GAP): Same architecture but replaces flatten with Global Average Pooling.
• ResNet‑18: A pre‑activated residual network fine‑tuned on each dataset.

4. Installation and Dependencies

Typical dependencies include:

• numpy
• opencv-python
• scikit-learn
• matplotlib
• tqdm
• torch
• torchvision
• jupyter

Clone this repository and install required packages:

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

Notebooks

Notebook	Description
iCubWorld1_CNN_Improved.ipynb	Trains and evaluates SimpleCNN and ResNet‑18 on iCubWorld1.0.
iCubWorld_Traditional_CV_Improved.ipynb	Implements BoVW + SVM pipeline on iCubWorld1.0 with improved settings.
Cifar_10_Tradition_CV.ipynb	Applies BoVW + SVM on CIFAR-10 and compares with CNN results.

Requirements

numpy>=1.18.5  
opencv-python>=4.5.0  
matplotlib>=3.2.2  
scikit-learn>=0.24.0  
torch>=1.7.0  
torchvision>=0.8.1  
tqdm>=4.50.0  
jupyter>=1.0.0  

5. Running the Code

1. Ensure iCubWorld1.0 data is in Dataset/iCubWorld1/train and Dataset/iCubWorld1/test.
2. Launch Jupyter Lab or Notebook:

jupyter notebook

3. Open and run notebooks in order:
   • iCubWorld1_CNN_Improved.ipynb
   • iCubWorld_Traditional_CV_Improved.ipynb
   • Cifar_10_Tradition_CV.ipynb
4. Each notebook handles preprocessing, training, evaluation, and visualization of results.

6. Evaluation

The following table summarizes classification accuracy (%) on the iCubWorld1.0 test subsets:

Method	Background	Categorization	Demonstrator	Robot
BoVW+SVM (tf–idf+L2)	93.0	31.08	21.59	11.47
BoVW+SVM (Stop-Words)	100.0	28.04	25.24	13.84
SimpleCNN (Flatten)	51.5	51.37	80.41	7.98
SimpleCNN (GAP)	53.0	58.02	88.51	16.46
ResNet-18	64.0	81.16	93.50	25.06

Accuracy on CIFAR-10 is also reported in the notebooks for direct comparison.

7. Results and Discussion

• CNNs outperform traditional CV pipelines on challenging subsets (Categorization, Demonstrator) due to learned spatial features.
• BoVW+SVM excels on Background bias but suffers on generalization (Robot subset).
• Global Average Pooling improves SimpleCNN stability and performance over flatten-based architecture.
• ResNet‑18 achieves closest performance to human baselines, demonstrating the benefit of residual learning.

8. Use of Generative AI Tools

• ChatGPT was used for code troubleshooting, documentation drafting, and conceptual discussion.
• All AI-suggested code was manually reviewed and validated before inclusion.

9. Author

• Kunwei Song (Student ID: 11537231)

10. References

• Fanello et al., iCubWorld: Friendly Robots Help Building Good Vision Data-Sets, IROS 2013.
• Krizhevsky et al., ImageNet Classification with Deep Convolutional Neural Networks, NIPS 2012.
• Lin et al., Network In Network, ICLR 2014.
• LeCun et al., Gradient-based Learning Applied to Document Recognition, IEEE 1998.
• CIFAR-10 Dataset: https://www.cs.toronto.edu/~kriz/cifar.html
• iCubWorld1.0 Dataset: http://robots.epfl.ch/icubworld1