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..
PointMAC: Meta-Learned Adaptation for Robust Test-Time Point Cloud Completion
Authors: Linlian Jiang, Rui Ma, Li Gu, Ziqiang Wang, Xinxin Zuo, Yang Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on synthetic, simulated, and real-world datasets demonstrate that Point MAC achieves state-of-the-art results by refining each sample individually to produce high-quality completions. To the best of our knowledge, this is the first work to apply meta-auxiliary test-time adaptation to point cloud completion. |
| Researcher Affiliation | Academia | 1Concordia University 2Jilin University 3Mila Quebec AI Institute 4Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, MOE, China EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Sample-Specific Test-Time Adaptation Algorithm 2: Meta-Auxiliary Training |
| Open Source Code | Yes | Answer: [Yes] Justification: We will release the code and data processing scripts upon publication. |
| Open Datasets | Yes | In this section, we evaluate our method on three types of datasets: purely synthetic datasets (Shape Net [10], PCN [4]), a high-fidelity simulated scanning dataset (MVP [38]), and a real-world scanned dataset (KITTI [11]). |
| Dataset Splits | Yes | Shape Net-55 provides 41,952 training and 10,518 testing shapes across 55 categories for category-agnostic evaluation. Shape Net-34 offers 46,765 training shapes and 5,705 testing shapes from 34 categories, split into 3,400 seen-class and 2,305 unseen-class samples for category-specific generalization. |
| Hardware Specification | Yes | All experiments are conducted on two NVIDIA V100 GPUs. |
| Software Dependencies | No | The paper does not explicitly state specific software versions for its implementation, such as Python or library versions. |
| Experiment Setup | Yes | We train the model for 250 epochs on the PCN [4] and Shape Net [10] datasets, and for 200 epochs on MVP [38]. The batch size is set to 40 for PCN, 32 for Shape Net, and 44 for MVP. During the joint training phase, we apply equal learning rates for the primary and auxiliary branches, with α = β = 2.5 10 5. In the meta-training and meta-testing stages, we perform 3 inner-loop gradient update steps to adapt the shared encoder using the auxiliary losses Lsmr aux and Lad aux. Optimization is carried out using Stochastic Gradient Descent (SGD) without momentum or weight decay. |