Probabilistic Interactive 3D Segmentation with Hierarchical Neural Processes

Authors: Jie Liu, Pan Zhou, Zehao Xiao, Jiayi Shen, Wenzhe Yin, Jan-Jakob Sonke, Efstratios Gavves

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on four 3D point cloud datasets demonstrate that NPISeg3D achieves superior segmentation performance with fewer clicks while providing reliable uncertainty estimations. Project Page: https://jliu4ai. github.io/NPISeg3D_projectpage/.
Researcher Affiliation Academia 1University of Amsterdam, Amsterdam, The Netherlands 2Singapore Management University, Singapore 3Netherlands Cancer Institute, Amsterdam, The Netherlands.
Pseudocode No The paper describes the methodology in detail but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No Project Page: https://jliu4ai. github.io/NPISeg3D_projectpage/. ... we will release our code, along with the improved annotation tool, to facilitate future research.
Open Datasets Yes We follow the dataset setup as in prior work (Yue et al., 2023) and train our model on the Scan Net V2-Train dataset (Dai et al., 2017). ... For evaluation, we consider two types of datasets: (1) In-domain dataset: Scan Net V2Val (Dai et al., 2017)... (2) Out-of-domain datasets: S3DIS (Armeni et al., 2016)... Replica (Straub et al., 2019)... and KITTI-360 (Liao et al., 2022)... Additionally, we further evaluate the part-level segmentation capability of our NPISeg3D and existing models on the Part Net dataset (Mo et al., 2019).
Dataset Splits Yes We follow the dataset setup as in prior work (Yue et al., 2023) and train our model on the Scan Net V2-Train dataset (Dai et al., 2017). For evaluation, we consider two types of datasets: (1) In-domain dataset: Scan Net V2-Val (Dai et al., 2017)...
Hardware Specification Yes Training is performed on a single Tesla A6000 GPU with a batch size of 5.
Software Dependencies No The paper mentions using Minkowski Res16UNet34C (Choy et al., 2019) backbone and Adam optimizer, but does not specify version numbers for any software dependencies like Python, PyTorch, CUDA, etc.
Experiment Setup Yes We train NPISeg3D end-to-end for 600 epochs using the Adam optimizer with an initial learning rate of 0.0005. The learning rate is reduced by a factor of 0.1 after 500 epochs to facilitate convergence. Training is performed on a single Tesla A6000 GPU with a batch size of 5. The KL loss coefficient λklin Eq. (10) is set to 0.005. For the segmentation loss Lseg, we use a combination of cross-entropy loss and dice loss, with coefficients of 1 and 2, respectively.