PromptGS Visual Prompting for Tiny Object Reconstruction in 3DGS Optimization

Xun Wang, Xutao Xue, Siyuan Li, Shayer Shabab Utsho, Kun Li, Mengqi Ji

Abstract:

Reconstructing tiny floating objects in large-scale 3D scene remains a fundamental challenge for 3D Gaussian Splatting (3DGS). These objects often receive insufficient point density and gradient supervision during training due to limited visibility and low image-space saliency, making them difficult to recover even
after prolonged optimization. We present PromptGS, a visual prompting framework that incorporates lightweight human input to guide the 3DGS optimization process. PromptGS fuses projected 2D error maps with user-specified spatial prompts to form a 3D attention field, which acts as an optimization
prior to guide Gaussian densification, adaptive resampling, and multiview selection. This mechanism directs training efforts toward regions with high semantic relevance but low point density, improving reconstruction in areas that are frequently overlooked. Furthermore, we design a Gaussian scoring function that ranks candidates based on their improvement potential, ensuring efficient resource allocation. Moreover, PromptGS achieves multiview consistent rendering of small objects, indicating that
their geometry and appearance are faithfully reconstructed in 3D space rather than approximated through view-dependent texture projection. Experiments on public benchmarks and challenging synthetic scenes demonstrate that PromptGS consistently outperforms existing methods in both visual fidelity and efficiency.