DCF-3DGS Distractor-Aware COLMAP-Free 3D Gaussian Splatting

WACV 2027
Namtae Kim1, Haebeom Jung1, Jaesik Park1
1Seoul National University, Republic of Korea
Paper
DCF-3DGS teaser: unordered input images are used to reconstruct a Gaussian scene and camera poses, with predicted masks and rendered views for dynamic distractors.

Abstract

Pose-free 3D Gaussian Splatting (3DGS) has demonstrated remarkable potential in reconstructing high-fidelity 3D scenes from unposed images. However, current frameworks degrade severely in real-world environments with dynamic distractors, as these transient elements violate the strict photometric consistency required for joint pose-scene optimization and often trap the model in suboptimal local minima.

We propose a novel framework that mitigates this ambiguity using a dual coarse-to-fine curriculum. First, we introduce a coarse-to-fine anchor-growing strategy, starting optimization with a sparse, low-frequency anchor grid to stabilize initial pose alignment and geometry. Second, we present a DINOv2-guided uncertainty-aware masking module to dynamically isolate transient objects.

The masking module is controlled by a coarse-to-fine uncertainty regularization schedule that initially enforces static-scene assumptions to stabilize poses before progressively allowing dynamic masking. Experiments show that our method reduces pose error among the compared methods in cluttered, unordered captures, while maintaining competitive Novel View Synthesis quality.

Overview of DCF-3DGS, showing coarse pose initialization and outlier filtering, coarse-to-fine anchor growth, DINOv2-guided uncertainty-aware optimization, and camera relocalization.
Overall pipeline. From unordered images, we extract a coarse structure and filter unreliable camera poses. The resulting geometry initializes anchor Gaussians that progressively increase in resolution. A DINO-guided MLP masks dynamic distractors under a coarse-to-fine regularization schedule, enabling robust joint pose-scene optimization. Finally, outlier cameras are relocalized using the optimized scene.

Video Comparisons

Mountain

Input Images
Input images for the Mountain scene
DCF-3DGS
T-3DGS

Patio High

Input Images
Input images for the Patio High scene
DCF-3DGS
T-3DGS

Spot

Input Images
Input images for the Spot scene
DCF-3DGS
T-3DGS

Image Comparisons

Fountain

DCF-3DGS reconstruction of the Fountain scene
DCF-3DGS
T-3DGS reconstruction of the Fountain scene
T-3DGS

Patio

DCF-3DGS reconstruction of the Patio scene
DCF-3DGS
T-3DGS reconstruction of the Patio scene
T-3DGS

Patio High

DCF-3DGS reconstruction of the Patio High scene
DCF-3DGS
AsymGS reconstruction of the Patio High scene
AsymGS

Spot

DCF-3DGS reconstruction of the Spot scene
DCF-3DGS
AsymGS reconstruction of the Spot scene
AsymGS

BibTeX


@inproceedings{Kim_DCF3DGS_2027,
  Title={Distractor-Aware COLMAP-Free 3D Gaussian Splatting},
  Author={Namtae Kim and Haebeom Jung and Jaesik Park},
  Booktitle={Proceedings of the Winter Conf. on Applications of Computer Vision (WACV)},
  Year={2027}
}