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..
OpenHype: Hyperbolic Embeddings for Hierarchical Open-Vocabulary Radiance Fields
Authors: Lisa Weijler, Sebastian Koch, Fabio Poiesi, Timo Ropinski, Pedro Hermosilla
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Open Hype both quantitatively and qualitatively on existing benchmarks, where we consistently outperform state-of-the-art methods. |
| Researcher Affiliation | Academia | Lisa Weijler TU Wien Sebastian Koch Ulm University Fabio Poiesi Fondazione Bruno Kessler Timo Ropinski Ulm University Pedro Hermosilla TU Wien |
| Pseudocode | No | The paper describes the methodology in prose and through figures, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| 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 will make the code publicly available upon acceptance. |
| Open Datasets | Yes | Dataset. We adapt the recent Search3D dataset [42] to the tasks of open-vocabulary segmentation and localization from radiance fields. This dataset provides instance segmentation masks at two levels of granularity: objects and object parts. The dataset provides annotations for two different sets of 3D scans, Multi Scan [27] and Scan Net++ [46]. ... We select the benchmark proposed by LERF [15] and later refined by Lang Splat [35]. ... A full list of assets used in this work, including the license, is given below. ... Scan Net++ subset of Search3D [42, 46] (https://github.com/aycatakmaz/search3d/tree/main/search3d/benchmark/docs/scannetpp_data_search3d): released under the original Scan Net++ data terms of use |
| Dataset Splits | Yes | As our test frames, we use the novel view synthesis split of Scan Net++. ... We work with undistorted images and sample a maximum of 250 train frames per scene evenly across all images. Note that while train and test frames are strictly separated for training and evaluating our models, ground truth poses are used for camera parameters to fairly compare against baselines. |
| Hardware Specification | Yes | Our method does not necessitate high-performance or specialized compute resources for training or inference; most experiments were conducted on a standard GPU setup using an NVIDIA RTX A5000. |
| Software Dependencies | Yes | We implement Open Hype in Nerfstudio [43] and build upon the Nerfacto model. ... Nerfstudio and its Nerfacto model [43] v1.1.5 (https://github.com/nerfstudio-project/nerfstudio): Apache License 2.0 Open Clip [5, 13] v2.29.0 (https://github.com/mlfoundations/open_clip): MIT License Semantic SAM [24] (https://github.com/UX-Decoder/Semantic-SAM) SAM [17] (https://github.com/facebookresearch/segment-anything): Apache License, Version 2.0, January 2004 |
| Experiment Setup | Yes | All auto-encoders are trained for 1000 epochs using Adam W [26] as optimizer with a weight decay value of 1e-4 and One Cycle Lr [41] as learning rate scheduler. The initial and final division factors are 10 and 1000, the percentage of the cycle (in number of steps) spent increasing the learning rate is 5%. We use a batch size of 10 images... We train all models for 30000 steps (approx. 60 min. per scene). |