Dr Relja Arandjelović (Др Реља Аранђеловић)

Links towards some of my projects and online demos. Most of these are links towards external pages .

Blog post: Towards Robust and Verified AI

Specification Testing, Robust Training, and Formal Verification

Code: On the Effectiveness of Interval Bound Propagation for Training Verifiably Robust Models

Training models that are verifiably robust to norm-bounded adversarial perturbations.

Project page and code: Neighbourhood Consensus Networks

Finding reliable dense correspondences between a pair of images, trained from weak supervision.

Blog post: Objects that Sound

What can be learnt by looking at and listening to a large number of unlabelled videos?

Dataset: Celebrity Together

194k images that have multiple labeled celebrities per image.

Project page and code: CNN Geometric matching with weak supervision

End-to-end weakly-supervised semantic alignment.

Project page and code: CNN Geometric matching

Convolutional neural network architecture for geometric matching.

Project page and code: NetVLAD

CNN architecture for weakly supervised place recognition.

Code: VGG Image Search Engine (VISE)

Image search engine mainly based on the codebase I developed during my PhD and PostDoc at VGG. It implements memory efficient large-scale retrieval using local descriptors, spatial reranking, Hamming embedding, query expansion, etc.

Offline demo: NetVLAD

Place recognition results using NetVLAD for all queries in the 24/7 Tokyo dataset.

Project page: Faces in Places

Faces in Places: Compound query retrieval.

Dataset: Celebrity in Places

36k images of celebrities in different types of scenes.

Project page: 24/7 Place Recognition by View Synthesis

Place recognition for situations where the scene undergoes a major change in appearance.

Project page: Sculpture Retrieval and Identification

A retrieval based method for automatically determining the title and sculptor of an imaged sculpture.

Demo: Smooth Sculpture Retrieval

Instantly search for specific smooth sculptures.

Click on the image, then outline the suggested (or any other) query region and click on 'Search'.

Project page: Visual Search of BBC News

On-the-fly retrieval of object categories, instances and faces using a textual keyword.

Demo: Visual Search of BBC News

Enter a text query in the textbox to the right, and hit search.

Example queries: Pound note, Mona Lisa, Buckingham palace, Big Ben, Coca Cola.

Further usage instructions
Blog post by the BBC Archive

Demo: Matching Ballad Illustrations

Instantly match and compare printed illustrations in the Bodleian library ballads.

Click on the image, then outline the suggested (or any other) query region and click on 'Search'.

Code: Fast Semantic Segmentation via Soft Segments

An implementation of the fast semantic segmentation method described in our ACCV 2014 paper Visual vocabulary with a semantic twist.