Skip to content

Latest commit

 

History

History
36 lines (22 loc) · 1.39 KB

File metadata and controls

36 lines (22 loc) · 1.39 KB

Image Similarity Matching

MLMI12 - Computer Vision Mini Project. ( Report )

Training a neural network to perform image similarity matching on the Tiny Imagenet dataset. Given two images identify whether they belong to the same class or different classes.

Methods

Multiclass Classification

Classify each image separately and check whether the classes match.

Multiclass Features with Cosine Similarity

Instead of getting the final class from the network, take the feature embeddings and calculate the cosine similarity between them. If the similarity is above a threshold, then the images are considered to be similar.

Siamese Network with Triplet Loss

Train a siamese network with triplet loss. The network takes in two images and outputs the feature embeddings. The triplet loss is calculated between the anchor, positive and negative images.

End-to-End Binary Classifier

Train a network end-to-end to provide binary classes for either similar or dissimal.

Results

Seen-Seen : Both image classes were seen during training Seen-Unseen : One image class was seen during training and the other was unseen Unseen-Unseen : Both image classes were unseen during training