Segment Anything in 3D with NeRFs

Authors: Jiazhong Cen, Zanwei Zhou, Jiemin Fang, chen yang, Wei Shen, Lingxi Xie, Dongsheng Jiang, XIAOPENG ZHANG, Qi Tian

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct various (e.g., object, part-level) segmentation tasks on the Replica [51] and NVOS [47] datasets. In this section, we quantitatively evaluate the segmentation ability of SA3D on various datasets. Then, we qualitatively demonstrate the versatility of SA3D, which can conduct instance segmentation, part segmentation, and text-prompted segmentation etc.
Researcher Affiliation Collaboration 1 Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Huawei Inc. 3 School of EIC, Huazhong University of Science and Technology
Pseudocode No The paper describes the method in prose and equations but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/Jumpat/Segment Anythingin3D.
Open Datasets Yes For quantitative experiments, we use the Neural Volumetric Object Selection (NVOS) [47], SPIn Ne RF [39], and Replica [51] datasets.
Dataset Splits No By default, we utilize all available views in the training set I. The views are uniformly sampled from the sorted training set.
Hardware Specification Yes On an Nvidia RTX 3090 GPU, the 3D segmentation process with 5 views can be completed within 10 seconds.
Software Dependencies No We implement SA3D using Py Torch [42] with reference to the code provided by DVGOv2 [54].
Experiment Setup Yes For our Ne RF model, we primarily employ the Tenso RF [3], utilizing the VM-48 representation to store the radiance latent vectors. The radiance fields are pre-trained for most datasets with 40,000 iterations. For the LLFF dataset [37] and the 360 dataset [1], the radiance fields are trained with 20,000 iterations.