Retri3D: 3D Neural Graphics Representation Retrieval

Authors: Yushi Guan, Daniel Kwan, Jean Dandurand, Xi Yan, Ruofan Liang, Yuxuan Zhang, Nilesh Jain, Nilesh Ahuja, Selvakumar Panneer, Nandita Vijaykumar

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that Retri3D is compatible with any NGR representation. On the LERF and Scan Net++ datasets, we show significant improvement in retrieval accuracy compared to existing techniques, while being orders of magnitude faster and storage efficient. [...] 4 EXPERIMENTS AND RESULTS
Researcher Affiliation Collaboration Yushi Guan1, Daniel Kwan1, Jean Sebastien Dandurand1, Xi Yan1, Ruofan Liang1, Yuxuan Zhang1, Nandita Vijaykumar1 [...] Nilesh Jain2, Nilesh Ahuja2, Selvakumar Panneer2 [...] 1 University of Toronto 2 Intel
Pseudocode No The paper describes methods through narrative text and figures but does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper states: "We adapt LERF and Lang Splat methods for scene retrieval because they are open-sourced", referring to baseline methods used, not the authors' own code for Retri3D. There is no explicit statement or link indicating that the source code for Retri3D is released or provided.
Open Datasets Yes We demonstrate the effectiveness of Retri3D using the LERF and Scan Net++ datasets (Kerr et al., 2023; Yeshwanth et al., 2023).
Dataset Splits No The paper mentions using 13 scenes from LERF and 280 scenes from Scan Net++ for evaluation. It describes metrics like P@k for retrieval accuracy, which implies evaluating against the entire set of available scenes for a query. However, it does not specify explicit training/validation/test splits for the Retri3D retrieval model itself.
Hardware Specification Yes Experiments are conducted on a desktop with an Intel i7-13700K CPU, Nvidia RTX 4090 GPU, and 64GB of RAM. We use Py Torch 2.1.2 with CUDA 12.0 on Ubuntu 22.04 LTS.
Software Dependencies Yes We use Py Torch 2.1.2 with CUDA 12.0 on Ubuntu 22.04 LTS.
Experiment Setup Yes All models are trained using the default configuration for 30,000 epochs. For Nerfacto, we also enable the "use-gradient-scaling" option to scale the gradient near the camera, reducing artifacts and creating a stronger baseline for rendering from random poses (Philip & Deschaintre, 2023).