Self-supervised deep learning methods for joint depth and ego-motion estimation can yield accurate trajectories without needing ground-truth training data. However, as they typically use photometric losses, their performance can degrade significantly when the assumptions these losses make (e.g. temporal illumination consistency, a static scene, and the absence of noise and occlusions) are violated. This limits their use for e.g. nighttime sequences, which tend to contain many point light sources (including on dynamic objects) and low signal-to-noise ratio (SNR) in darker image regions. In this paper, we show how to use a combination of three techniques to allow the existing photometric losses to work for both day and nighttime images. First, we introduce a per-pixel neural intensity transformation to compensate for the light changes that occur between successive frames. Second, we predict a per-pixel residual flow map that we use to correct the reprojection correspondences induced by the estimated ego-motion and depth from the networks. And third, we denoise the training images to improve the robustness and accuracy of our approach. These changes allow us to train a single model for both day and nighttime images without needing separate encoders or extra feature networks like existing methods. We perform extensive experiments and ablation studies on the challenging Oxford RobotCar dataset to demonstrate the efficacy of our approach for both day and nighttime sequences.
We present RangeUDF, a new implicit representation based framework to recover the geometry and semantics of continuous 3D scene surfaces from point clouds. Unlike occupancy fields or signed distance fields which can only model closed 3D surfaces, our approach is not restricted to any type of topology. Being different from the existing unsigned distance fields, our framework does not suffer from any surface ambiguity. In addition, our RangeUDF can jointly estimate precise semantics for continuous surfaces. The key to our approach is a range-aware unsigned distance function together with a surface-oriented semantic segmentation module. Extensive experiments show that RangeUDF clearly surpasses state-of-the-art approaches for surface reconstruction on four point cloud datasets. Moreover, RangeUDF demonstrates superior generalization capability across multiple unseen datasets, which is nearly impossible for all existing approaches.
Sampling is a key operation in point-cloud task and acts to increase computational efficiency and tractability by discarding redundant points. Universal sampling algorithms (e.g., Farthest Point Sampling) work without modification across different tasks, models, and datasets, but by their very nature are agnostic about the downstream task/model. As such, they have no implicit knowledge about which points would be best to keep and which to reject. Recent work has shown how task-specific point cloud sampling (e.g., SampleNet) can be used to outperform traditional sampling approaches by learning which points are more informative. However, these learnable samplers face two inherent issues: i) overfitting to a model rather than a task, and \ii) requiring training of the sampling network from scratch, in addition to the task network, somewhat countering the original objective of down-sampling to increase efficiency. In this work, we propose an almost-universal sampler, in our quest for a sampler that can learn to preserve the most useful points for a particular task, yet be inexpensive to adapt to different tasks, models, or datasets. We first demonstrate how training over multiple models for the same task (e.g., shape reconstruction) significantly outperforms the vanilla SampleNet in terms of accuracy by not overfitting the sample network to a particular task network. Second, we show how we can train an almost-universal meta-sampler across multiple tasks. This meta-sampler can then be rapidly fine-tuned when applied to different datasets, networks, or even different tasks, thus amortizing the initial cost of training.
Scene flow is a powerful tool for capturing the motion field of 3D point clouds. However, it is difficult to directly apply flow-based models to dynamic point cloud classification since the unstructured points make it hard or even impossible to efficiently and effectively trace point-wise correspondences. To capture 3D motions without explicitly tracking correspondences, we propose a kinematics-inspired neural network (Kinet) by generalizing the kinematic concept of ST-surfaces to the feature space. By unrolling the normal solver of ST-surfaces in the feature space, Kinet implicitly encodes feature-level dynamics and gains advantages from the use of mature backbones for static point cloud processing. With only minor changes in network structures and low computing overhead, it is painless to jointly train and deploy our framework with a given static model. Experiments on NvGesture, SHREC'17, MSRAction-3D, and NTU-RGBD demonstrate its efficacy in performance, efficiency in both the number of parameters and computational complexity, as well as its versatility to various static backbones. Noticeably, Kinet achieves the accuracy of 93.27% on MSRAction-3D with only 3.20M parameters and 10.35G FLOPS.
Unmanned aerial vehicles (UAVs) have been used for many applications in recent years, from urban search and rescue, to agricultural surveying, to autonomous underground mine exploration. However, deploying UAVs in tight, indoor spaces, especially close to humans, remains a challenge. One solution, when limited payload is required, is to use micro-UAVs, which pose less risk to humans and typically cost less to replace after a crash. However, micro-UAVs can only carry a limited sensor suite, e.g. a monocular camera instead of a stereo pair or LiDAR, complicating tasks like dense mapping and markerless multi-person 3D human pose estimation, which are needed to operate in tight environments around people. Monocular approaches to such tasks exist, and dense monocular mapping approaches have been successfully deployed for UAV applications. However, despite many recent works on both marker-based and markerless multi-UAV single-person motion capture, markerless single-camera multi-person 3D human pose estimation remains a much earlier-stage technology, and we are not aware of existing attempts to deploy it in an aerial context. In this paper, we present what is thus, to our knowledge, the first system to perform simultaneous mapping and multi-person 3D human pose estimation from a monocular camera mounted on a single UAV. In particular, we show how to loosely couple state-of-the-art monocular depth estimation and monocular 3D human pose estimation approaches to reconstruct a hybrid map of a populated indoor scene in real time. We validate our component-level design choices via extensive experiments on the large-scale ScanNet and GTA-IM datasets. To evaluate our system-level performance, we also construct a new Oxford Hybrid Mapping dataset of populated indoor scenes.
With the recent availability and affordability of commercial depth sensors and 3D scanners, an increasing number of 3D (i.e., RGBD, point cloud) datasets have been publicized to facilitate research in 3D computer vision. However, existing datasets either cover relatively small areas or have limited semantic annotations. Fine-grained understanding of urban-scale 3D scenes is still in its infancy. In this paper, we introduce SensatUrban, an urban-scale UAV photogrammetry point cloud dataset consisting of nearly three billion points collected from three UK cities, covering 7.6 km^2. Each point in the dataset has been labelled with fine-grained semantic annotations, resulting in a dataset that is three times the size of the previous existing largest photogrammetric point cloud dataset. In addition to the more commonly encountered categories such as road and vegetation, urban-level categories including rail, bridge, and river are also included in our dataset. Based on this dataset, we further build a benchmark to evaluate the performance of state-of-the-art segmentation algorithms. In particular, we provide a comprehensive analysis and identify several key challenges limiting urban-scale point cloud understanding. The dataset is available at http://point-cloud-analysis.cs.ox.ac.uk.
Ubiquitous positioning for pedestrian in adverse environment has served a long standing challenge. Despite dramatic progress made by Deep Learning, multi-sensor deep odometry systems yet pose a high computational cost and suffer from cumulative drifting errors over time. Thanks to the increasing computational power of edge devices, we propose a novel ubiquitous positioning solution by integrating state-of-the-art deep odometry models on edge with an EKF (Extended Kalman Filter)-LoRa backend. We carefully compare and select three sensor modalities, i.e., an Inertial Measurement Unit (IMU), a millimetre-wave (mmWave) radar, and a thermal infrared camera, and realise their deep odometry inference engines which runs in real-time. A pipeline of deploying deep odometry considering accuracy, complexity, and edge platform is proposed. We design a LoRa link for positional data backhaul and projecting aggregated positions of deep odometry into the global frame. We find that a simple EKF based fusion module is sufficient for generic positioning calibration with over 34% accuracy gains against any standalone deep odometry system. Extensive tests in different environments validate the efficiency and efficacy of our proposed positioning system.
Direction finding and positioning systems based on RF signals are significantly impacted by multipath propagation, particularly in indoor environments. Existing algorithms (e.g MUSIC) perform poorly in resolving Angle of Arrival (AoA) in the presence of multipath or when operating in a weak signal regime. We note that digitally sampled RF frontends allow for the easy analysis of signals, and their delayed components. Low-cost Software-Defined Radio (SDR) modules enable Channel State Information (CSI) extraction across a wide spectrum, motivating the design of an enhanced Angle-of-Arrival (AoA) solution. We propose a Deep Learning approach to deriving AoA from a single snapshot of the SDR multichannel data. We compare and contrast deep-learning based angle classification and regression models, to estimate up to two AoAs accurately. We have implemented the inference engines on different platforms to extract AoAs in real-time, demonstrating the computational tractability of our approach. To demonstrate the utility of our approach we have collected IQ (In-phase and Quadrature components) samples from a four-element Universal Linear Array (ULA) in various Light-of-Sight (LOS) and Non-Line-of-Sight (NLOS) environments, and published the dataset. Our proposed method demonstrates excellent reliability in determining number of impinging signals and realized mean absolute AoA errors less than $2^{\circ}$.
Camera localization is a fundamental and crucial problem for many robotic applications. In recent years, using deep-learning for camera-based localization has become a popular research direction. However, they lack robustness to large domain shifts, which can be caused by seasonal or illumination changes between training and testing data sets. Data augmentation is an attractive approach to tackle this problem, as it does not require additional data to be provided. However, existing augmentation methods blindly perturb all pixels and therefore cannot achieve satisfactory performance. To overcome this issue, we proposed RADA, a system whose aim is to concentrate on perturbing the geometrically informative parts of the image. As a result, it learns to generate minimal image perturbations that are still capable of perplexing the network. We show that when these examples are utilized as augmentation, it greatly improves robustness. We show that our method outperforms previous augmentation techniques and achieves up to two times higher accuracy than the SOTA localization models (e.g., AtLoc and MapNet) when tested on `unseen' challenging weather conditions.