Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians
Authors: Xiyu Zhang, Chong Bao, YiPeng Chen, Hongjia Zhai, Yitong Dong, Hujun Bao, Zhaopeng Cui, Guofeng Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method outperforms state-of-the-art approaches in both indoor and urban scenes, delivering superior surface reconstruction quality. |
| Researcher Affiliation | Academia | Xiyu Zhang1 Chong Bao1 Yipeng Chen1 Hongjia Zhai1 Yitong Dong1 Hujun Bao1 Zhaopeng Cui1 Guofeng Zhang1 1State Key Lab of CAD & CG, Zhejiang University |
| Pseudocode | No | The paper describes methods in prose and with figures (e.g., Figure 2: Overview of Atlas GS), but does not contain any clearly labeled 'Pseudocode' or 'Algorithm' blocks or sections. |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We use official datasets in experiments, and code will be released. |
| Open Datasets | Yes | We evaluate our method using well-known datasets for both indoor and outdoor scene reconstruction. For indoor environments, we use Scan Net [50], Scan Net++ [51], and Replica [52]. For outdoor settings, we employ the Matrix City [58] dataset for surface reconstruction. |
| Dataset Splits | Yes | In line with previous studies [13, 14], we select four scenes from Scan Net, seven from Replica, and four from Scan Net++, sampling images uniformly from the sequences. For outdoor evaluation, four city blocks from Matrix City are used. Additional details are provided in the supplementary material. ... Similar to previous works for indoor scene reconstruction [14], we select four scenes in Scan Net [50], including scene0050_00, scene0084_00, scene0580_00, scene0616_00 and seven scenes in Replica [52], office0~office3, room0~room2, and as for Scan Net++ [51], we select four scenes, 8b5caf3398, b20a261fdf, f34d532901, f6659a3107. ... For each scene in Scan Net [50] and Replica [52], we select one out of every 10 images in the original image sequence. For Scan Net++ [51], we use the image sequence from the i Phone and select one out of every 60 images. All the images are cropped and resized, and center-cropped to 640 480. For Matrix City [58], we use all the provided images and make the image resolution 960 540. |
| Hardware Specification | Yes | All experiments are performed on a single NVIDIA 4090D GPU. |
| Software Dependencies | No | Our implementation is based on Py Torch [55], incorporating a custom surfel rasterization module for semantic learning, and optimize the parameters with Adam optimizer [56]. The paper mentions software tools like PyTorch and Adam optimizer, but does not provide specific version numbers for them or any other libraries. |
| Experiment Setup | Yes | Implementation Details. We implement our approach in Py Torch [55], incorporating a custom surfel rasterization module for semantic learning, and optimize the parameters with Adam optimizer [56]. The hyperparameter K is set to 10, and the voxel size is fixed at 0.01. During training, the loss weights λ1, λ2, λ3, λ4, λ5, λ6 are set to 0.25, 0.1, 0.1, 1.0, 100, and 0.05, respectively. The explicit plane indicator is initially derived from the semantic lifted Sf M points and is reinitialized using the semantic Gaussians if the discrepancy exceeds a predefined threshold. Surfaces are extracted using TSDF Fusion [57]. All experiments are performed on a single NVIDIA 4090D GPU. Additional implementation details are provided in the supplementary material. ... Most of the training learning rates are similar to those used in [54]. We set the hyperparameter K to 10 for indoor scenes and 5 for urban scenes, with a voxel size of 0.01, and the feature dim is 32 in our sparse feature grid. For all scenes, the implicit-structured Gaussian is trained for 40,000 steps. Voxels grow between steps 1,500 and 20,000, provided the gradients of the Gaussians exceed 2e-4 and are pruned if the opacities of all local Gaussians fall below 0.005. During training, we start our 3D global planar regularization from step 7000 and 2D local surface regularization from 20000. |