We propose a novel end-to-end method for cross-view pose estimation. Given a ground-level query image and an aerial image that covers the query's local neighborhood, the 3 Degrees-of-Freedom camera pose of the query is estimated by matching its image descriptor to descriptors of local regions within the aerial image. The orientation-aware descriptors are obtained by using a translationally equivariant convolutional ground image encoder and contrastive learning. The Localization Decoder produces a dense probability distribution in a coarse-to-fine manner with a novel Localization Matching Upsampling module. A smaller Orientation Decoder produces a vector field to condition the orientation estimate on the localization. Our method is validated on the VIGOR and KITTI datasets, where it surpasses the state-of-the-art baseline by 72% and 36% in median localization error for comparable orientation estimation accuracy. The predicted probability distribution can represent localization ambiguity, and enables rejecting possible erroneous predictions. Without re-training, the model can infer on ground images with different field of views and utilize orientation priors if available. On the Oxford RobotCar dataset, our method can reliably estimate the ego-vehicle's pose over time, achieving a median localization error under 1 m and a median orientation error of around 1$^{\circ }$at 14 FPS.

Objects in videos are typically characterized by continuous smooth motion. We exploit continuous smooth motion in three ways. 1) Improved accuracy by using object motion as an additional source of supervision, which we obtain by anticipating object locations from a static keyframe. 2) Improved efficiency by only doing the expensive feature computations on a small subset of all frames. Because neighboring video frames are often redundant, we only compute features for a single static keyframe and predict object locations in subsequent frames. 3) Reduced annotation cost, where we only annotate the keyframe and use smooth pseudo-motion between keyframes. We demonstrate computational efficiency, annotation efficiency, and improved mean average precision compared to the state-of-the-art on four datasets: ImageNet VID, EPIC KITCHENS-55, YouTube-BoundingBoxes and Waymo Open dataset. Our source code is available at https://github.com/L-KID/Video-object-detection-by-location-anticipation.

In this paper, we propose a simple attention mechanism, we call Box-Attention. It enables spatial interaction between grid features, as sampled from boxes of interest, and improves the learning capability of transformers for several vision tasks. Specifically, we present BoxeR, short for Box Transformer, which attends to a set of boxes by predicting their transformation from a reference window on an input feature map. The BoxeR computes attention weights on these boxes by considering its grid structure. Notably, BoxeR-2D naturally reasons about box information within its attention module, making it suitable for end-to-end instance detection and segmentation tasks. By learning invariance to rotation in the box-attention module, BoxeR-3D is capable of generating discriminative information from a bird's-eye view plane for 3D end-to-end object detection. Our experiments demonstrate that the proposed BoxeR-2D achieves state-of-the-art results on COCO detection and instance segmentation. Besides, BoxeR-3D improves over the end-to-end 3D object detection baseline and already obtains a compelling performance for the vehicle category of Waymo Open, without any class-specific optimization. Code is available at https://github.com/kienduynguyen/BoxeR.

Not all video frames are equally informative for recognizing an action. It is computationally infeasible to train deep networks on all video frames when actions develop over hundreds of frames. A common heuristic is uniformly sampling a small number of video frames and using these to recognize the action. Instead, here we propose full video action recognition and consider all video frames. To make this computational tractable, we first cluster all frame activations along the temporal dimension based on their similarity with respect to the classification task, and then temporally aggregate the frames in the clusters into a smaller number of representations. Our method is end-to-end trainable and computationally efficient as it relies on temporally localized clustering in combination with fast Hamming distances in feature space. We evaluate on UCF101, HMDB51, Breakfast, and Something-Something V1 and V2, where we compare favorably to existing heuristic frame sampling methods.

This work proposes a metric learning approach for self-supervised scene flow estimation. Scene flow estimation is the task of estimating 3D flow vectors for consecutive 3D point clouds. Such flow vectors are fruitful, e.g. for recognizing actions, or avoiding collisions. Training a neural network via supervised learning for scene flow is impractical, as this requires manual annotations for each 3D point at each new timestamp for each scene. To that end, we seek for a self-supervised approach, where a network learns a latent metric to distinguish between points translated by flow estimations and the target point cloud. Our adversarial metric learning includes a multi-scale triplet loss on sequences of two-point clouds as well as a cycle consistency loss. Furthermore, we outline a benchmark for self-supervised scene flow estimation: the Scene Flow Sandbox. The benchmark consists of five datasets designed to study individual aspects of flow estimation in progressive order of complexity, from a moving object to real-world scenes. Experimental evaluation on the benchmark shows that our approach obtains state-of-the-art self-supervised scene flow results, outperforming recent neighbor-based approaches. We use our proposed benchmark to expose shortcomings and draw insights on various training setups. We find that our setup captures motion coherence and preserves local geometries. Dealing with occlusions, on the other hand, is still an open challenge.

During simultaneous localization and mapping, geometrical constraints are established between map features. These constraints, introduced through measurements and motion prediction, produce a bending effect in the event of closing a large loop. In this paper we present a discussion of the bending problem for trajectory based representations. Furthermore, we propose a generic approach to reduce the bending effect that exploits common geometrical constraints in human-made environments through the use of a heading based curvature measurement. We show by means of experimental results that our approach increases significantly both the global and local map accuracy.

An annotated data set is presented meant to help researchers in developing, evaluating, and comparing various approaches in robotics for building space representations appropriate for communicating with humans. The data consist of omnidirectional images, laser range scans, sonar readings, and robot odometry. A set of base-level human spatial concepts is used to annotate the data.

Vision systems are used more and more in 'personal' robots interacting with humans, since semantic information about objects and places can be derived from the rich sensory information. Visual information is also used for building appearance based topological maps, which can be used for localization. In this paper we describe a system capable of using this appearance based topological map for navigation. The system is made robust by using the epipolar geometry and a planar floor constraint in computing the necessary heading information. Using this method the robot is able to drive robustly in a large environment. We tested the method on real data under varying environment conditions and compared performance with a human-controlled robot.

In appearance based robot localization a new image is matched with every image in the database. In this paper we describe how to reduce the number of images in this database with minimal loss of information and thereby increasing the efficiency of localization significantly. First we build an appearance based model consisting of a graph in which the nodes denote images and links denote relations between images. This graph is then pruned using the connected dominating set algorithm. The method is tested on an image set acquired by a mobile robot equipped with an omnidirectional camera driving around in an office environment, as well as sets of images taken while driving through a real furnished home environment

An ambitious goal in modern robotic science is to build mobile robots that are able to interact as companions in real world environments. Especially for caretaking of elderly people a system robustly working at private homes is essential, requiring a very natural and human oriented way of communication. Since home environments are usually very individual a first task for a newly acquired robot is to get familiar with its new environment. This paper gives a short overview on how we integrated a vision based localization using the advantages of a very modular architecture and extending a spoken dialog system for online labeling and interaction about different locations. We present results from the integrated system working in a real, fully furnished home environment where it was able to learn the names of different rooms. This system enables us to perform real user studies in future without the need to fall back to Wizard-of-Oz experiments. Ongoing work aims at enabling the robot to take initiative by asking for unknown locations. A future extension is the ability to generalize over features of known rooms to make predictions when encountering unknown rooms