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..
D$^2$GS: Dense Depth Regularization for LiDAR-free Urban Scene Reconstruction
Authors: Kejing Xia, Jidong Jia, Ke Jin, Yucai BAI, Li Sun, Dacheng Tao, Youjian Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on the Waymo dataset demonstrate that our method consistently outperforms stateof-the-art methods, producing more accurate geometry even when compared with those using ground-truth Li DAR data. (...) We conducted a detailed experimental analysis using the Waymo dataset [37], a widely-used dataset for real-world driving scenarios. (...) We adopt PSNR, SSIM, and LPIPS as metrics to evaluate the quality of image reconstruction. (...) We compare our method against several Gaussian Splatting approaches: 3DGS [5], S3GS [7], PVG [8], and Omni Re [1], and Ne RF-based approach Emer Ne RF [27]. (...) We perform ablation studies on a Waymo subset (12 sequences, detailed in supplementary material) to assess the impact of D2GS s key components: Progressive Pruning (P.P.), Depth Enhancer (D.E.), and Road Node (R.N.). |
| Researcher Affiliation | Collaboration | 1Wuhan University 2Shanghai Jiaotong University 3Tong Ji University 4Bosch 5Nanyang Technological University |
| Pseudocode | No | No explicitly labeled pseudocode or algorithm blocks are present. The paper describes methods through structured text and mathematical formulations but does not present them in a pseudocode format. |
| Open Source Code | No | The code will be open-sourced after the company s review process. |
| Open Datasets | Yes | We conducted a detailed experimental analysis using the Waymo dataset [37], a widely-used dataset for real-world driving scenarios. |
| Dataset Splits | Yes | We selected the challenging Waymo dataset subset, NOTR Dynamic32, proposed by [27]. We utilize data from the three frontal cameras (FRONT, FRONT RIGHT, FRONT LEFT), resizing the images to a resolution of 640 960. (...) Our main experiments are conducted on the Waymo NOTR Dynamic32 [27] dataset, comparing our method against S3GS [7], PVG [8], Omni Re [1], and Li DAR-free baselines. The specific segment IDs are listed in Tab. 5. Our ablation studies are performed on a 12-sequence subset of the Waymo NOTR dataset, with segment IDs provided in Tab. 6. |
| Hardware Specification | Yes | All other methods were evaluated by us using their official codebases on a single NVIDIA H20 GPU. |
| Software Dependencies | No | Our pipeline begins by generating initial depth estimates using Depth Splat [21] (weight name: "depthsplat-gs-base-dl3dv-256x448-randview2-6-d94d996f.pth") , which processes input images and camera parameters to produce per-view depth maps. |
| Experiment Setup | Yes | The model contains 60,000 iterations of training, divided into 40,000 iterations for RGB-only training and 20,000 iterations for iterative Gaussian and depth enhancement training. (...) Our proposed Road Node s learning rate is 0.05. The Depth Enhancer module is activated halfway through training, with the diffusion process consisting of 80 steps and an early stop mechanism that prevents the smoothness loss from increasing for 8 continuous steps. (...) Table 4: Value of the key hyperparameters in our method. Hyperparameter Value Description λC 2.0 Confidence threshold in Depth Enhancer λref 1.0 Weight for reference depth loss in Depth Enhancer λsmooth 1 8 Weight for depth smoothness loss in Depth Enhancer λw 1 16 Weight for warping loss in Depth Enhancer N 20 Depth Enhancer update frequency λnormal 10.0 Weight for road normal alignment loss in Road Node λflat 1.0 Weight for road flatness loss in Road Node λroad 0.1 Weight for Road Node loss |