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..
RPG360: Robust 360 Depth Estimation with Perspective Foundation Models and Graph Optimization
Authors: Dongki Jung, Jaehoon Choi, Yonghan Lee, Dinesh Manocha
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on benchmark datasets Matterport3D [5], Stanford2D3D [4], and 360Loc [26] using images of a resolution 1024 512. For Matterport3D and Stanford2D3D, we adopt the official train and test splits. We additionally evaluate on Matterport3D-2K (2048 1024) for high-resolution benchmarks, following [47]. 360Loc consists of four scenes, including both indoor and outdoor environments, each providing panoramic sequences of mapping and query images. We employ all mapping sequences from 360Loc, which include ground truth poses and depth maps, to evaluate the zero-shot performance of depth estimation. As monocular depth estimation increasingly emphasizes 3D structural awareness for practical applications [52, 40], we adopt 3D metrics, such as Chamfer Distance, F-Score and Io U, rather than 2D metrics to better assess improvements in 3D structure and geometry. Further details of the 3D metrics are described in the supplementary materials. |
| Researcher Affiliation | Academia | Dongki Jung Jaehoon Choi Yonghan Lee Dinesh Manocha University of Maryland, College Park EMAIL |
| Pseudocode | No | The paper describes the approach using text and diagrams (Figure 3), but no explicit pseudocode or algorithm blocks are provided. |
| Open Source Code | No | Answer: [No] Justification: We use the public datasets and cite them in the paper. |
| Open Datasets | Yes | We conduct extensive experiments on benchmark datasets Matterport3D [5], Stanford2D3D [4], and 360Loc [26] using images of a resolution 1024 512. |
| Dataset Splits | Yes | For Matterport3D and Stanford2D3D, we adopt the official train and test splits. |
| Hardware Specification | Yes | We use the Adam optimizer [33] to perform gradient descent on a single RTX A5000 GPU. |
| Software Dependencies | No | The paper mentions using the Adam optimizer, but does not provide specific version numbers for any software libraries, frameworks, or programming languages (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | We use the Adam optimizer [33] to perform gradient descent on a single RTX A5000 GPU. To accelerate convergence, we adopt a multi-scale approach following [50]. The ERP depth map DERP l R h/2l w/2l where l {0, ..., L 1}, is downsampled by a factor of 2l. In this experiment, we set L = 3 and use learning rates of 5 10l L. Each optimization step is performed for 300, 150, and 30 iterations. The weights of the loss terms ηp, ηd, and ηn are set to 50, 0.5, and 10, respectively. |