Assignment1_GNR638

Scene Recognition with Bag-of-Words: UC Merced Land Use Dataset

The UC Merced Land Use Dataset contains 21 classes of aerial images, with 100 images per class. Each image has a resolution of 256 × 256 pixels.

Dataset Split

Training Set: 70% of data per class.

Validation Set: 10% of data per class.

Test Set: 20% of data per class.

Overview

The goal of this project is to perform classical computer vision topic, image recognition. In particular, examining the task of scene recognition beginning with simplest method, tiny images and KNN(K nearest neighbor) classification, and then move forward to the state-of-the-art techniques, bags of quantized local features and linear classifiers learned by SVC(support vector classifier).

Reference : "Scene-recognition-with-bag-of-words"

Implementation steps

1. Dataset Pre-processing
2. Training
3. Optimization of codewords
4. Visualizing
5. Accuracy

Steps To Run

• Use dataset "split_dataset_new_new (1)"
• Install cyvlfeat
• Run "final3.py"

Results

1. TSNE_Visualization
2. SVM_Confusion_Matrix
3. NN_Confusion_Matrix
4. Accuracy vs No. of Codeworks plot
5. Predicted_label vs True_label Confusion_Matrix

Contributors

• Priya Nemani
• Rita Mahato