LEG-SLAM: Real-Time Language-Enhanced Gaussian Splatting for SLAM

Overview

LEG-SLAM is an open-vocabulary 3D SLAM system that integrates 3D Gaussian Splatting, DINOv2 feature extraction, and Talk2DINO language grounding to enable real-time semantic 3D scene understanding. Unlike existing methods, LEG-SLAM allows text-driven interactive exploration of reconstructed environments without predefined object categories.

🔹 Key Features:

• Real-time 3D Reconstruction: High-fidelity scene reconstruction with Gaussian Splatting.
• Open-Vocabulary Understanding: Uses DINOv2 features and Talk2DINO to match text queries to visual features.
• Efficient Feature Compression: PCA-based embedding compression enables low-latency inference.
• Interactive Scene Queries: Retrieve semantic masks in real-time by specifying objects via text.
• High-Speed Performance: Achieves 10 FPS on Replica and 18 FPS on ScanNet, significantly faster than prior methods.

🔬 Methodology

1. DINOv2 Feature Extraction: Extracts rich, self-supervised embeddings from RGB frames.
2. Talk2DINO Language Alignment: Transforms text queries into DINOv2-compatible feature space.
3. PCA Compression: Reduces embeddings from 768D → 64D, enabling real-time processing.
4. 3D Gaussian Splatting: Constructs a continuous, high-resolution 3D scene representation.
5. Semantic Querying: Computes cosine similarity between scene embeddings and textual queries, generating semantic heatmaps.

📜 Paper

If you find this work useful, please cite:

@article{LEG-SLAM,
  title={LEG-SLAM: Open-Vocabulary 3D Gaussian Splatting for SLAM},
  author={Anonymous Authors},
  journal={Under Review at ICCV 2025},
  year={2025}
}

⏳ Code Availability

🔹 Code will be released upon paper acceptance. Stay tuned for updates! 🚀