Invited Talks

Title: A unifying view of geometry, semantics, and data association in SLAM

Abstract: Recent years have seen impressive progress in robot perception, including accurate visual-inertial odometry, dense metric reconstruction, and object recognition in real time. Existing approaches to simultaneous localization and mapping (SLAM), however, rely on low-level geometric features and do not take advantage of object-level information. This talk will focus on a unifying view of geometry, semantics, and data association in SLAM. A major contribution of our approach is the use of structured object models to build meaningful maps online and probabilistic data association that avoids making hard, potentially wrong associations between semantic observations and objects in ambiguous environments.

Title: Multiple Target Tracking from UAVs using the Recursive-RANSAC Algorithm

Abstract: In this talk we will describe the newly developed Recursive-RANSAC algorithm for multiple target tracking, with an emphasis on applications to tracking ground based targets from fixed-wing and multi-rotor UAVs.

Title: Higher-order Projected Power Iterations for Scalable Multi-Matching

Abstract: The matching of multiple objects is a fundamental problem in vision, robotics and graphics. In order to robustly handle ambiguities, noise and repetitive patterns in challenging real-world settings, it is essential to take geometric consistency between points into account. In this talk the Higher-order Projected Power Iteration (HiPPI) algorithm will be presented, which is (i) efficient and scales to tens of thousands of points, (ii) straightforward to implement (only few lines of code), (iii) able to incorporate geometric consistency, (iv) guarantees cycle-consistent multi-matchings, and (iv) comes with theoretical convergence guarantees.

Title: Towards certifiably robust spatial perception

Abstract: Spatial perception is concerned with the estimation of a world model –that describes the state of the robot     and the environment– using sensor data and prior knowledge. As such, it includes a broad set of robotics  and computer vision problems, ranging from object detection and pose estimation to robot localization and mapping. Most perception algorithms require extensive and application-dependent parameter tuning and often fail in off-nominal conditions (e.g., in the presence of large noise, outliers, and incorrect data association). In this talk, I will present recent advances in the design of certifiably robust spatial perception algorithms that are robust to extreme amounts of outliers and afford performance guarantees. I show these algorithms can achieve unprecedented robustness in a variety of applications, ranging from mesh registration and image-based object localization, to SLAM.

Title: Learning representations for data association 

Abstract: We formulate data association as a graph embedding problem then use a Graph Convolutional Network to learn an appropriate embedding function for aligning image features. We use cycle consistency to train our network in an unsupervised fashion, since ground truth correspondence is difficult or expensive to acquire. In addition, geometric consistency losses can be added at training time, even if the information is not available in the test set, unlike previous approaches that optimize cycle consistency directly.

Title: Consistent multi-view data association  

Abstract: Many robotics applications require alignment and fusion of observations obtained at multiple views to form a global map of the environment. Multi-way data association methods provide a mechanism to improve alignment accuracy of pairwise associations and ensure their consistency. However, existing methods that solve this computationally challenging problem are often too slow for real-time applications. In this talk, we will present a recent technique based on leveraging insights from the spectral graph clustering literature to obtain cycle consistent and accurate associations in a computationally efficient manner. This algorithmic framework can provide significant improvement in the accuracy and efficiency of existing discrete assignment problems, which traditionally use pairwise (but potentially inconsistent) correspondences.

Title: Learning motion and place descriptor from LiDARs for long-term navigation

Abstract: This talk introduces two different deep learning-based approaches namely for place recognition and ego-motion estimation. The first part of the talk presents a robust place descriptor, "scan context". Combining this descriptor with a simple CNN allowed us to memorize places for the global localization while maintaining the generalizability. The performance was evaluated using the monthly obtained year-long NCLT dataset. By training from a single day data and testing over a year data, we verified the algorithm works for the long-term place recognition. The next topic includes unsupervised deep learning-based LiDAR odometry. We combined two types of losses for the learning, ) Uncertainty-weighted ICP Loss and FOV loss. When training, we used pre-trained weights from KITTI and applied to other datasets to secure generalization capability. 

Title: A Research Agenda for Robust Semantic SLAM

Abstract: This talk will discuss some of the challenges and opportunities in developing robust semantic mapping and localization algorithms, which we believe is essential for creating long-lived autonomous mobile robots that can operate safely with minimal human supervision in highly dynamic and/or unstructured environments.

Title: Learning Representations from Monocular Vision for Robust Navigation

Abstract: -

Title: Distributed consistent matching via clustering

Abstract: Data matching, the problem of associating identities of different data points while ensuring logical consistency relations, is a fundamental problem in robotics. The task becomes even more relevant and challenging in multi-agent settings, where the data is typically spread at different locations, and communications can be, from an energy point of view, quite expensive. In this talk, we first review the QuickMatch algorithm for multi-image feature matching, which ensures consistency constraints by construction using clustering. We then present a scheme for distributing the matching across computational units (agents) that largely preserves feature match quality and minimizes communication between agents (avoiding, in particular, the need of flooding all data to all agents).

Title: Learning Correspondences for 3D Reconstruction and Pose Estimation 

Abstract: 3D reconstruction and pose estimation are two fundamental problems in 3D computer vision. In most approaches to these problems, the foremost challenge is to establish correspondences between observations and 3D models. In this talk, I would like to discuss how to make use of learning-based methods to solve the correspondence problems in 3D reconstruction and pose estimation. To illustrate, I will introduce some of our recent works in this direction. The first is a transformation-invariant feature descriptor based on group CNNs with applications in SfM and visual localization. The second is a pixel-wise voting network to robustly establish 3D-2D correspondences for 6DoF pose estimation.  The final part of this talk introduces a multi-view matching method for multi-view 3D human pose estimation in crowd scenes.