Fine-Grained Image Localisation

Final assignment for Computer Vision (COMP90086). Given a dataset of images with their associated location label, we aim to investigate methods to recognize the location of an unlabelled image. To this end, we benchmarked various feature decriptors/extractors and image matching techniques.

Descriptors/Features:

• SIFT
• ASIFT
• NetVLAD (Original Code)
• Self-supervised CNN (Rotation)
• Self-supervised CNN (Warping)

Image Matching:

• KNN
• FLANN
• MLP

Dependencies

This program was developed using:

1. Python 3.9.7
2. Keras 2.6.0
3. Tensorflow 2.6.0

Pretrained weights for NetVLAD "netvlad_weights.h5" must be downloaded and put into the checkpoint directory before running. Download here (Obtained from Github repository "Netvlad-Keras")

How to Run the Code

The main file is "Experiments.ipynb" After running, the extracted features are stored in the checkpoint directory. Additonally, 'predictions_netvlad_knn.csv' is the final prediction output used for the Kaggle competition (link) where the predictions were ranked 6th.