GLS: Geometry-aware 3D Language Gaussian Splatting

Abstract: Recently, 3D Gaussian Splatting (3DGS) has achieved impressive performance on indoor surface reconstruction and 3D open-vocabulary segmentation. This paper presents GLS, a unified framework of 3D surface reconstruction and open-vocabulary segmentation based on 3DGS. GLS extends two fields by improving their sharpness and smoothness. For indoor surface reconstruction, we introduce surface normal prior as a geometric cue to guide the rendered normal, and use the normal error to optimize the rendered depth. For 3D open-vocabulary segmentation, we employ 2D CLIP features to guide instance features and enhance the surface smoothness, then utilize DEVA masks to maintain their view consistency. Extensive experiments demonstrate the effectiveness of jointly optimizing surface reconstruction and 3D open-vocabulary segmentation, where GLS surpasses state-of-the-art approaches of each task on MuSHRoom, ScanNet++ and LERF-OVS datasets. Project webpage: https://jiaxiongq.github.io/GLS_ProjectPage.

• The paper presents a novel joint optimization strategy that integrates surface normal priors with semantic cues to improve 3D surface reconstruction.
• It leverages 2D CLIP features for robust open-vocabulary segmentation, achieving superior results on benchmarks like MuSHRoom, ScanNet++, and LERF-OVS.
• The GLS framework introduces innovative geometric and semantic regularization terms that enhance training efficiency and rendering quality.

An Overview of "GLS: Geometry-aware 3D Language Gaussian Splatting"

The paper entitled "GLS: Geometry-aware 3D Language Gaussian Splatting" explores the intersection of surface reconstruction and open-vocabulary segmentation utilizing 3D Gaussian Splatting (3DGS) within a unified framework. Authored by Jiaxiong Qiu, Liu Liu, Zhizhong Su, and Tianwei Lin from Horizon Robotics, the paper introduces GLS, a novel approach that leverages the inherent correlations between these two tasks to enhance performance significantly.

Key Contributions

The primary contributions of GLS are centered on a joint optimization strategy that integrates both geometric and semantic cues for improved 3D surface reconstruction and segmentation. Key innovations include:

1. Integration of Surface Normal Priors: Leveraging geometric cues via surface normal priors enables the GLS framework to optimize rendered depth through normal error minimization, thus enhancing surface reconstruction accuracy.
2. Semantic Feature Enhancement: By employing 2D CLIP features, GLS enhances the consistency of instance features across different views. This is particularly beneficial in open-vocabulary segmentation tasks where view-inconsistency can challenge feature integrity.
3. Two Novel Regularization Terms: GLS introduces novel geometric and semantic regularizations. Specifically, the framework includes a smoothness enhancement for rendered depth, guided by Gaussian semantic features, and a depth refinement mechanism based on normal error guidance.
4. Improved Framework Efficiency: The study demonstrates that incorporating semantic attributes into the 3DGS framework not only retains but enhances efficiency during training and rendering operations.

Experimental Validation

The utility and robustness of GLS are documented through extensive experimentation on diverse datasets, including MuSHRoom, ScanNet++, and LERF-OVS. The proposed method outperforms existing state-of-the-art techniques in both tasks:

• Surface Reconstruction: GLS surpasses approaches such as 2DGS and PGSR, demonstrating superior performance in established metrics like Chamfer L1 distance and Normal Consistency. The integration of semantic masks also facilitates the reduction of reconstruction noise induced by challenging lighting conditions.
• Open-Vocabulary Segmentation: When compared with models like OpenGaussian and Gaussian Grouping, GLS yields higher fidelity in segmentation tasks as evidenced by mIoU and mBIoU metrics. It successfully mitigates the view-inconsistent noise commonly plaguing 2D-driven segmentation frameworks.

Implications and Future Directions

The fusion of geometric and semantic cues within GLS has significant implications. Practically, it enhances the applicability of real-time 3DGS in complex indoor environments, addressing challenges in augmented and virtual reality scenarios. Theoretically, it underscores the benefits of task interdependency, suggesting future models could further exploit intertwined characteristics of adjacent computational tasks.

The paper posits that future developments in AI could focus on broadening the adaptability of model-agnostic frameworks like GLS, especially by potentially incorporating more complex or diverse geometric and semantic priors. The adaptation of similar methodologies could guide advancements in other 3D representation and processing tasks, thereby improving the robustness and versatility of computer vision algorithms in dynamic scenes.

In conclusion, "GLS: Geometry-aware 3D Language Gaussian Splatting" presents a comprehensive and technically progressive approach that enriches the functionalities of 3D Gaussian Splatting, proposing a versatile solution that blends novel theoretical insights with cutting-edge practical implementations for real-time interactive environments.