[New] The final version of the HPatches dataset is now available




Introduction


Local features are at the centre of many fundamental computer vision problems. While local features have been subject of study in computer vision for almost twenty years, recent significant progress in this area has led to substantial improvements in many computer vision tasks including registration, stereo vision, motion estimation, matching, retrieval, recognition of objects and actions. With the advent of modern deep learning techniques, an area of particular interest is the development of learnable local feature descriptors. While this direction has already proved to be quite fruitful, there are still many challenges left for a variety of problem classes, including defining appropriate training sets, evaluation protocols, and benchmarks. This workshop will promote both a technical discussion on methods to construct and learn better local features as well as new approaches to a rigorous and realistic evaluation of their performance. Accepted papers and extended abstracts will be presented as posters / spotlights at the workshop, in areas including, but not limited to:

  • Review of the state of the art and recent developments.
  • Local features: open problems.
  • Deep learning and local features.
  • Review of datasets for learning and evaluation of descriptors.
  • Introduction of novel datasets and benchmark protocols.
  • Methods to make deep detectors/descriptors practical. Pushing the limits in terms of size, convolutions, binary output etc.

We also invite authors to submit results to our descriptor evaluation challenge. The final results will be presented at the workshop.

Important Dates

Paper Submission Deadline: 26 Aug
Paper Acceptance Notification: 5 Sep
Camera Ready Date: 9 Sep
Challenge Submission Deadline: 3 Oct
Extended Abstract Submission Deadline: 2 Sep
Workshop Date: 10 Oct

Venue

The workshop is part of the ECCV 2016 workshops. Please see the ECCV webpage for details on venue, accomodations, and other details!


Challenge

We provide a newly collected dataset of local patches (HPatches) extracted from image sequences.
This dataset is accompanied by a benchmarking suite (HBench) that evaluates the performance of feature descriptors.

images images-easy patches-easy

Researchers are encouraged to download the dataset and report their descriptor's performance using HBench. Submissions have to be made to the organizers before 3rd of October 2016, in order to be considered for the final presentation of the results during the conference.

HPatches dataset https://github.com/featw/hpatches
HBench benchmarking suite https://github.com/featw/hbench

Please submit your results for the challenge using this link

Programme

  • 08:30 - 09:30 / Progress and state of the art review
    • Feature detectors - Andrea Vedaldi
    • Feature descriptors - Krystian Mikolajczyk
  • 09:30 - 11:50 / Recent developments
    • 09:30-10:10 / Learning to compare image patches - Sergey Zagoruyko
    • 10:10-10:30 / Coffee break
    • 10:30-11:10 / Metric learning techniques for convolutional patch descriptors and segmentation-aware neural networks - Iasonas Kokkinos
    • 11:10-11:50 / LIFT: Learned Invariant Feature Transform - Vincent Lepetit
  • 11:50 - 12:30 / Introduction of new dataset and challenge results
    • Datasets and methods - Vassileios Balntas
    • Challenge and results - Karel Lenc
  • 12:30 - 14:00 / Posters
    • 12:30-13:00 / Poster spotlight
    • 13:00-14:00 / Poster session


Poster presentations
The CUDA LATCH Binary Descriptor: Because Sometimes Faster Means Better Christopher Parker, Matthew Daiter, Kareem Omar, Gil Levi and Tal Hassner
Sensor Fusion for Sparse SLAM with Descriptor Pooling Philipp Tiefenbacher, Julian Heuser, Timo Schulze, Mohammadreza Babaee and Gerhard Rigoll
An Evaluation of Local Feature Detectors and Descriptors for Infrared Images Johan Johansson, Atsuto Maki and Martin Solli
Evaluating Local Features for Day-Night Matching Hao Zhou, Torsten Sattler, David Jacobs
Image Correspondences Matching Using Multiple Features Fusion Song Wu and Michael Lew
Learning Local Convolutional Features for Face Recognition with 2D-Warping Harald Hanselmann and Hermann Ney
Improving Performances of MSER Features in Matching and Retrieval Tasks Andrzej Sluzek

Keynote speakers

Iasonas Kokkinos , Ecole Central Paris

Vincent Lepetit, Technical University Graz

Stefano Soatto , UCLA

For full programme, and keynote abstracts please click here

People

Jiri Matas, Czech Technical University

Krystian Mikolajczyk, Imperial College London

Tinne Tuytelaars, KU Leuven

Andrea Vedaldi, University of Oxford

Vassileios Balntas, Imperial College London

Karel Lenc, University of Oxford