Semi-supervised learning (SSL) has promising potential for improving model performance using both labelled and unlabelled data. Since recovering 3D information from 2D images is an ill-posed problem, the current state-of-the-art methods of monocular object detection (Mono3D) have relatively low precision recall, making semi-supervised Mono3D tasks challenging understudied. In this work, we propose a unified effective framework called Mix-Teaching that can be applied to most detectors. Based...

This paper proposes a method to extract the position and pose of vehicles in 3D world from single traffic camera. Most previous monocular vehicle detection algorithms focused on cameras perspective driver, assumed known intrinsic extrinsic calibration. On contrary, this focuses same task using uncalibrated cameras. We observe that homography between road plane image is essential data synthesis for task, can be estimated without camera intrinsics extrinsics. conduct by estimating rotated...

We propose a novel approach for monocular 3D object detection by leveraging local perspective effects of each object. While the global effect shown as size and position variations has been exploited extensively, perspectives long overlooked. design module to regress newly defined variable named keyedge-ratios parameterization shape distortion account perspective, derive depth yaw angle from it. Theoretically, this does not rely on pixel-wise or in image objects, therefore independent camera...

The main goal of this paper is to introduce the data collection effort at Mcity targeting automated vehicle development. We captured a comprehensive set from perception sensors (Lidars, Radars, Cameras) as well steering/brake/throttle inputs and an RTK unit. Two in-cabin cameras record human driver's behaviors for possible future use. naturalistic driving on selected open roads recorded different time day weather conditions. also perform designed choreography inside test facility focusing...

This paper proposes a correspondence-free method for point cloud rotational registration. We learn an embedding each in feature space that preserves the SO(3)-equivariance property, enabled by recent developments equivariant neural networks. The proposed shape registration achieves three major advantages through combining learning with implicit models. First, necessity of data association is removed because permutation-invariant property network architectures similar to PointNet. Second, can...

This article reports on recent progress in robot perception and control methods developed by taking the symmetry of problem into account. Inspired existing mathematical tools for studying structures geometric spaces, sensor registration, state estimator, provide indispensable insights formulations generalization robotics algorithms to challenging unknown environments. When combined with computational learning hard-to-measure quantities, symmetry-preserving unleash tremendous performance. The...

Object detection and tracking is a key task in autonomy. Specifically, 3D object have been an emerging hot topic recently. Although various methods proposed for detection, uncertainty the tasks has less explored. Uncertainty helps us tackle error perception system improve robustness. In this paper, we propose method improving target performance by adding regression to SECOND detector, which one of most representative algorithms detection. Our estimates positional dimensional uncertainties...

Autonomous driving (AV) has been intensively researched over the last decade. In this paper, we introduce an open autonomous software stack to enable faster design, development, and testing of algorithms on simulated or experimental vehicles, which hope will become a useful tool for AV researchers.

This paper reports a new continuous 3D loss function for learning depth from monocular images. The dense prediction image is supervised using sparse LIDAR points, which enables us to leverage available open source datasets with camera-LIDAR sensor suites during training. Currently, accurate and affordable range not readily available. Stereo cameras LIDARs measure either inaccurately or sparsely/costly. In contrast the current point-to-point evaluation approach, proposed treats point clouds...

We propose a framework to edit real-world Lidar scans with novel object layouts while preserving realistic background environment. Compared the synthetic data generation frameworks where point clouds are generated from scratch, our focuses on new scenario in given environment, and method also provides labels for data. This approach ensures remains relevant specific aiding both development evaluation of algorithms scenarios. view synthesis, allows creation counterfactual scenarios significant...

A unified neural network structure is presented for joint 3D object detection and point cloud segmentation in this paper. We leverage rich supervision from both labels rather than using just one of them. In addition, an extension based on single-stage detectors proposed the implicit function widely used scene understanding. The branch takes final feature map module as input, produces that generates semantic distribution each its corresponding voxel center. demonstrated performance our...

This paper reports a new continuous 3D loss function for learning depth from monocular images. The dense prediction image is supervised using sparse LIDAR points, which enables us to leverage available open source datasets with camera-LIDAR sensor suites during training. Currently, accurate and affordable range not readily available. Stereo cameras LIDARs measure either inaccurately or sparsely/costly. In contrast the current point-to-point evaluation approach, proposed treats point clouds...

3D object detection with a single image is an essential and challenging task for autonomous driving. Recently, keypoint-based monocular has made tremendous progress achieved great speed-accuracy trade-off. However, there still exists huge gap LIDAR-based methods in terms of accuracy. To improve their performance without sacrificing efficiency, we propose sort lightweight feature pyramid network called Lite-FPN to achieve multi-scale fusion effective efficient way, which can boost the...

This paper proposes an equivariant neural network that takes data in any semi-simple Lie algebra as input. The corresponding group acts on the adjoint operations, making our proposed adjoint-equivariant. Our framework generalizes Vector Neurons, a simple $\mathrm{SO}(3)$-equivariant network, from 3-D Euclidean space to spaces, building upon invariance property of Killing form. Furthermore, we propose novel bracket layers and geometric channel mixing extend modeling capacity. Experiments are...