| CARVIEW |
Second Landmark Recognition Workshop
This workshop fosters research on image retrieval and landmark recognition by introducing a novel large-scale dataset, together with evaluation protocols. Here is a Google AI blog detailing the workshop and the challenge. Our 1st workshop is held in CVPR 2018.
Recognition Challenge
https://www.kaggle.com/c/landmark-recognition-2019Label famous (and not-so-famous) landmarks in image.
Retrieval Challenge
https://www.kaggle.com/c/landmark-retrieval-2019Given an image, can you find all of the same landmarks in a dataset?
Challenge Timeline
Challenge Begins
Challenge is held on Kaggle
Apr. 8 2019Final Submission Deadline
Jun. 3 2019
Workshop
Held with CVPR’19
Jun. 16 2019, 8:30 - 12:30, PDTWorkshop Schedule
Invited Talk 1
Noah Snavely (Cornell Tech / Google AI)
End-to-End Geometric Learning
Coffee Break
10:00 - 10:30Recognition Challenge Winner Presentations
JL , GLRunner , smlyaka (Hosted by Bohyung Han) 10:30 - 11:10Retrieval Challenge Winner Presentations
smlyaka , imagesearch , Layer 6 AI (Hosted by Xu Zhang) 11:10 - 11:50Invited Talk 2
Krystian Mikolajczyk (Imperial College London)
Recent Progress and Remaining Challenges in Local Features
Close Remark
12:25 - 12:35Invited Speakers
Noah Snavely
Associate Professor at Cornell Tech
End-to-End Geometric Learning
One often hears that vision systems should be trained end-to-end using deep learning. This talk will explore this idea in the context of 3D geometry, presenting end-to-end methods for a number of tasks, including keypoint detection, pose estimation, and view synthesis. I will show how the use of geometric reasoning as an end goal of learning can enable emergent discovery of good keypoints, systems for predicting 3D shape from single images, and more, all without the use of explicit supervision. I will relate these ideas back to the landmark recognition problem.
Krystian Mikolajczyk
Associate Professor at Imperial College London
Recent Progress and Remaining Challenges in Local Features
In many computer vision applications local image features and descriptors have been replaced by end-to-end learning based methods but still remain the preferred choice for estimating accurate 3D models in multiple view geometry, camera and object pose estimation, or efficient SLAM. Large benchmarks and synthetically generated training data stimulated the progress in nearly all areas of computer vision including keypoint detection and description. I will present some recent works in this domain including new multi-task benchmarks for feature matching and localization, new methods for keypoint extraction and description, as well as for improving their efficiency.