A-TDOM: Active TDOM via On-the-Fly 3DGS

TL;DR

A-TDOM addresses the latency of traditional TDOM generation by enabling near real-time, active TDOM construction through On-the-Fly SfM and incremental 3D Gaussian Splatting updates. It introduces a Gaussian Sampling and Integration scheme guided by reprojected Delaunay triangulations and adaptive optimization, plus orthographic splatting to render updated TDOMs after each image. Key contributions include (i) online base construction and per-image Gaussian insertion, (ii) region-aware training to avoid poorly constrained areas, and (iii) orthographic projection for accurate TDOM rendering with efficient updates. The framework demonstrates strong rendering quality and substantial gains in speed and memory efficiency on UAV and other large-scale datasets, making near real-time TDOM generation practical for production workflows.

Abstract

True Digital Orthophoto Map (TDOM), a 2D objective representation of the Earth's surface, is an essential geospatial product widely used in urban management, city planning, land surveying, and related applications. However, traditional TDOM generation typically relies on a complex offline photogrammetric pipeline, leading to substantial latency and making it unsuitable for time-critical or real-time scenarios. Moreover, the quality of TDOM may deteriorate due to inaccurate camera poses, imperfect Digital Surface Model (DSM), and incorrect occlusions detection. To address these challenges, this work introduces A-TDOM, a near real-time TDOM generation method built upon On-the-Fly 3DGS (3D Gaussian Splatting) optimization. As each incoming image arrives, its pose and sparse point cloud are computed via On-the-Fly SfM. Newly observed regions are then incrementally reconstructed as additional 3D Gaussians are inserted using a Delaunay triangulated Gaussian sampling and integration and are further optimized via adaptive training iterations and learning rate, especially in previously unseen or coarsely modeled areas. With orthogonal splatting integrated into the rendering pipeline, A-TDOM can actively produce updated TDOM outputs immediately after each 3DGS update. Code is now available at https://github.com/xywjohn/A-TDOM.

Figures (12)

• Figure 1: Our A-TDOM. For each input image, it pose and corresponding new point cloud are estimated and the current 3DGS field is updated via the proposed Gaussians Sampling and Integration method, followed by the 3DGS field optimization. Subsequently, TDOM is generated using the Orthogonal Splatting technique.
• Figure 2: The workflow of our A-TDOM. After an initial 3DGS field, for each new image, a Delaunay triangulation is applied on the 3D-2D reprojections to create a mask for 3DGS field updating and optimization. The TDOM is generated after each 3DGS field update via orthogonal splatting.
• Figure 3: Scene reconstruction based on Delaunay triangular masking. A Delaunay Triangulation is constructed on the image plane based on the 3D-2D reprojections, and the regions covered by the triangulation are defined as the key regions. During subsequent training, only key regions are used for the rendering loss.
• Figure 4: Gaussians Sampling and Integration. Previously unseen and coarsely reconstructed regions are identified based on gradient discrepancy map, and the extraction mask are formed. Subsequently, we sample points for Gaussians integration.
• Figure 5: Projection Transformation Process. In perspective projection, the Gaussians are projected from the frustum, whereas in orthographic projection, they are projected from the cuboid, both into the Normalized Device Coordinates (NDC) system.