In this letter, we present a neural field-based real-time monocular mapping framework for accurate and dense Simultaneous Localization and Mapping (SLAM). Recent neural mapping frameworks show promising results, but rely on RGB-D or pose inputs, or cannot run in real-time. To address these limitations, our approach integrates dense-SLAM with neural implicit fields. Specifically, our dense SLAM approach runs parallel tracking and global optimization, while a neural field-based map is constructed incrementally based on the latest SLAM estimates. For the efficient construction of neural fields, we employ multi-resolution grid encoding and signed distance function (SDF) representation. This allows us to keep the map always up-to-date and adapt instantly to global updates via loop closing. For global consistency, we propose an efficient $Sim(3)$ -based pose graph bundle adjustment (PGBA) approach to run online loop closing and mitigate the pose and scale drift. To enhance depth accuracy further, we incorporate learned monocular depth priors. We propose a novel joint depth and scale adjustment (JDSA) module to solve the scale ambiguity inherent in depth priors. Extensive evaluations across synthetic and real-world datasets validate that our approach outperforms existing methods in accuracy and map completeness while preserving real-time performance.

Urban environments are regions in which spectral variability and spatial variability are extremely high, with a huge range of shapes and sizes, and they also demand high resolution images for applications involving their study. Due to the fact that these environments can grow even more over time, applications related to their monitoring tend to turn to autonomous intelligent systems, which together with remote sensing data could help or even predict daily life situations. The task of mapping cities by autonomous operators was usually carried out by aerial optical images due to its scale and resolution; however new scientific questions have arisen, and this has led research into a new era of highly-detailed data extraction. For many years, using artificial neural models to solve complex problems such as automatic image classification was commonplace, owing much of their popularity to their ability to adapt to complex situations without needing human intervention. In spite of that, their popularity declined in the mid-2000s, mostly due to the complex and time-consuming nature of their methods and workflows. However, newer neural network architectures have brought back the interest in their application for autonomous classifiers, especially for image classification purposes. Convolutional Neural Networks (CNN) have been a trend for pixel-wise image segmentation, showing flexibility when detecting and classifying any kind of object, even in situations where humans failed to perceive differences, such as in city scenarios. In this paper, we aim to explore and experiment with state-of-the-art technologies to semantically label 3D urban models over complex scenarios. To achieve these goals, we split the problem into two main processing lines: first, how to correctly label the façade features in the 2D domain, where a supervised CNN is used to segment ground-based façade images into six feature classes, roof, window, wall, door, balcony and shop; second, a Structure-from-Motion (SfM) and Multi-View-Stereo (MVS) workflow is used to extract the geometry of the façade, wherein the segmented images in the previous stage are then used to label the generated mesh by a “reverse” ray-tracing technique. This paper demonstrates that the proposed methodology is robust in complex scenarios. The façade feature inferences have reached up to 93% accuracy over most of the datasets used. Although it still presents some deficiencies in unknown architectural styles and needs some improvements to be made regarding 3D-labeling, we present a consistent and simple methodology to handle the problem.

We present HI-SLAM2, a geometry-aware Gaussian SLAM system that achieves fast and accurate monocular scene reconstruction using only RGB input. Existing Neural SLAM or 3DGS-based SLAM methods often trade off between rendering quality and geometry accuracy, our research demonstrates that both can be achieved simultaneously with RGB input alone. The key idea of our approach is to enhance the ability for geometry estimation by combining easy-to-obtain monocular priors with learning-based dense SLAM, and then using 3D Gaussian splatting as our core map representation to efficiently model the scene. Upon loop closure, our method ensures on-the-fly global consistency through efficient pose graph bundle adjustment and instant map updates by explicitly deforming the 3D Gaussian units based on anchored keyframe updates. Furthermore, we introduce a grid-based scale alignment strategy to maintain improved scale consistency in prior depths for finer depth details. Through extensive experiments on Replica, ScanNet, Waymo Open, ETH3D SLAM and ScanNet++ datasets, we demonstrate significant improvements over existing Neural SLAM methods and even surpass RGB-D-based methods in both reconstruction and rendering quality.

The multi-lens multispectral cameras (MSCs), such as Micasense Rededge and Parrot Sequoia, can record multispectral information by each separated lenses. With their lightweight and small size, which making they are more suitable for mounting on an Unmanned Aerial System (UAS) to collect high spatial images for vegetation investigation. However, due to the multi-sensor geometry of multi-lens structure induces significant band misregistration effects in original image, performing band co-registration is necessary in order to obtain accurate spectral information. A robust and adaptive band-to-band image transform (RABBIT) is proposed to perform band co-registration of multi-lens MSCs. First is to obtain the camera rig information from camera system calibration, and utilizes the calibrated results for performing image transformation and lens distortion correction. Since the calibration uncertainty leads to different amount of systematic errors, the last step is to optimize the results in order to acquire a better co-registration accuracy. Due to the potential issues of parallax that will cause significant band misregistration effects when images are closer to the targets, four datasets thus acquired from Rededge and Sequoia were applied to evaluate the performance of RABBIT, including aerial and close-range imagery. From the results of aerial images, it shows that RABBIT can achieve sub-pixel accuracy level that is suitable for the band co-registration purpose of any multi-lens MSC. In addition, the results of close-range images also has same performance, if we focus on the band co-registration on specific target for 3D modelling, or when the target has equal distance to the camera.