dConstruct Robotics is proud to announce that our latest research, developed in collaboration with the National University of Singapore, has been selected as a Highlight Paper at the International Conference on Computer Vision (ICCV 2025).

Titled “A Constrained Optimisation Approach for Gaussian Splatting from Coarsely-posed Images and Noisy LiDAR Point Clouds,” the paper was recognised among the top 9.7% of accepted papers and the top 2.3% of all submissions. Out of 11,239 submissions this year, only 2,374 were accepted, making this distinction especially significant.

Advancing Real-World 3D Reconstruction

Our research tackles a fundamental barrier to scalable 3D reconstruction: achieving high-quality, metrically accurate outputs with speed and cost efficiency. We present a constrained optimisation framework designed to perform reliably with coarsely estimated camera poses and noisy LiDAR inputs, reducing dependence on computationally intensive preprocessing workflows. Our approach maintains and in some cases surpasses the performance of established 3DGS-COLMAP baselines.

This innovation is already integrated into dConstruct suite of products, including d.ASH Pack series and d.ASH Xplorer.

Overcoming Limitations of Existing Pipelines

In multi-camera SLAM pipelines, both camera poses and LiDAR point clouds are often noisy or only coarsely estimated. When used directly, such inputs can lead to blurred reconstructions and degraded geometry - limiting commercial viability, particularly in precision-critical use cases.

Traditional Structure-from-Motion (SfM) pipelines are typically used to refine camera parameters, but they are computationally intensive and time-consuming, creating bottlenecks for large-scale or time-sensitive applications.

Leveraging the NVIDIA RTX 6000 Ada Generation GPU and CUDA framework, our approach addresses these challenges directly - enabling faster, more efficient reconstruction pipelines while maintaining high geometric accuracy.

Read the full paper HERE

‍

‍

‍

‍

‍

‍