GLACE: Global Local Accelerated Coordinate Encoding **Authors:** Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, Marc Pollefeys **Affiliations:** Department of Computer Science, ETH Zurich; Microsoft Mixed Reality & AI Zurich Lab **Abstract:** Scene coordinate regression (SCR) methods are effective for small-scale scenes but struggle in large-scale scenes without ground truth 3D point clouds. This paper introduces GLACE, a novel method that integrates pre-trained global and local encodings to enable SCR to scale to large scenes using a single small-sized network. GLACE addresses the challenge of implicit triangulation in large scenes by proposing a feature diffusion technique that implicitly groups reprojection constraints with co-visibility, avoiding overfitting to trivial solutions. Additionally, a positional decoder parameterizes output positions more effectively for large-scale scenes. GLACE achieves state-of-the-art results on large-scale datasets without using 3D models or depth maps for supervision. **Contributions:** 1. GLACE is the first SCR method to achieve state-of-the-art performance on large-scale scenes without ensemble networks or 3D model supervision. 2. A novel feature diffusion technique integrates global and local encodings, effectively grouping reprojection constraints with co-visibility. 3. An improved positional decoder parameterizes output positions more effectively for large-scale scenes. **Related Work:** - **Pose Regression:** Methods encode the scene into a neural network and regress poses from query images. - **Feature Matching Based Localization:** Methods represent the 3D scene using 3D geometry and match pixels in query images to 3D points in the 3D model. **Experiments:** - **Datasets:** 7 Scenes, 12 Scenes, Cambridge Landmarks, Aachen Day-Night. - **Implementation:** PyTorch-based architecture, similar to ACE, with adjustments for large outdoor scenes. - **Evaluation:** GLACE outperforms state-of-the-art SCR methods on large-scale scenes, achieving comparable performance to feature matching methods with a smaller model size. **Ablation Study:** - **Feature Diffusion:** Enhances performance by effectively grouping reprojection constraints. - **Decoder:** Improves performance by allowing the model to parameterize a multimodal distribution. **Conclusion:** GLACE is a novel SCR method that scales to large scenes using a single network, leveraging co-visibility information and an improved position decoder.GLACE: Global Local Accelerated Coordinate Encoding **Authors:** Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, Marc Pollefeys **Affiliations:** Department of Computer Science, ETH Zurich; Microsoft Mixed Reality & AI Zurich Lab **Abstract:** Scene coordinate regression (SCR) methods are effective for small-scale scenes but struggle in large-scale scenes without ground truth 3D point clouds. This paper introduces GLACE, a novel method that integrates pre-trained global and local encodings to enable SCR to scale to large scenes using a single small-sized network. GLACE addresses the challenge of implicit triangulation in large scenes by proposing a feature diffusion technique that implicitly groups reprojection constraints with co-visibility, avoiding overfitting to trivial solutions. Additionally, a positional decoder parameterizes output positions more effectively for large-scale scenes. GLACE achieves state-of-the-art results on large-scale datasets without using 3D models or depth maps for supervision. **Contributions:** 1. GLACE is the first SCR method to achieve state-of-the-art performance on large-scale scenes without ensemble networks or 3D model supervision. 2. A novel feature diffusion technique integrates global and local encodings, effectively grouping reprojection constraints with co-visibility. 3. An improved positional decoder parameterizes output positions more effectively for large-scale scenes. **Related Work:** - **Pose Regression:** Methods encode the scene into a neural network and regress poses from query images. - **Feature Matching Based Localization:** Methods represent the 3D scene using 3D geometry and match pixels in query images to 3D points in the 3D model. **Experiments:** - **Datasets:** 7 Scenes, 12 Scenes, Cambridge Landmarks, Aachen Day-Night. - **Implementation:** PyTorch-based architecture, similar to ACE, with adjustments for large outdoor scenes. - **Evaluation:** GLACE outperforms state-of-the-art SCR methods on large-scale scenes, achieving comparable performance to feature matching methods with a smaller model size. **Ablation Study:** - **Feature Diffusion:** Enhances performance by effectively grouping reprojection constraints. - **Decoder:** Improves performance by allowing the model to parameterize a multimodal distribution. **Conclusion:** GLACE is a novel SCR method that scales to large scenes using a single network, leveraging co-visibility information and an improved position decoder.