Adaptive Multi-step Refinement Network for Robust Point Cloud Registration

Authors: Zhi Chen, Yufan Ren, Tong Zhang, Zheng Dang, Wenbing Tao, Sabine Susstrunk, Mathieu Salzmann

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments demonstrate our method s effectiveness; we achieve state-of-the-art recall rates on the 3DMatch/3DLo Match and KITTI benchmarks. Notably, on 3DLo Match, our method reaches 80.4% recall rate, with an absolute improvement of 1.2%. Our code is available at https://github.com/Zhi Chen902/AMR.
Researcher Affiliation Academia Zhi Chen EMAIL HUST Yufan Ren EMAIL IVRL, EPFL Tong Zhang EMAIL IVRL, EPFL Zheng Dang EMAIL CVLab, EPFL Wenbing Tao EMAIL HUST Sabine Süsstrunk EMAIL IVRL, EPFL Mathieu Salzmann EMAIL CVLab, EPFL
Pseudocode No The paper describes methods and processes using mathematical formulas and descriptive text, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present any code-like formatted procedures.
Open Source Code Yes Our code is available at https://github.com/Zhi Chen902/AMR.
Open Datasets Yes Following previous research (Yu et al., 2021; Qin et al., 2022; Yang et al., 2022), we evaluate our method using the 3DMatch (Zeng et al., 2017) and 3DLo Match (Huang et al., 2021) benchmarks. These two datasets are created from 62 RGB-D scenes, with 46 scenes used for training, 8 for validation, and 8 for testing. KITTI odometry (Geiger et al., 2012) consists of 11 sequences of driving scenes scanned using Li DAR. Following the methodology of previous studies (Qin et al., 2022; Choy et al., 2019), we use sequences 0-5 for training, 6-7 for validation, and 8-10 for testing.
Dataset Splits Yes These two datasets are created from 62 RGB-D scenes, with 46 scenes used for training, 8 for validation, and 8 for testing. Following the methodology of previous studies (Qin et al., 2022; Choy et al., 2019), we use sequences 0-5 for training, 6-7 for validation, and 8-10 for testing.
Hardware Specification Yes We train the KITTI model for 80 epochs and the 3DMatch/3DLo Match models for 20 epochs, taking around 48 hours on a single V100 GPU.
Software Dependencies No The paper mentions using KP-Conv and Geo Transformer as backbone architectures and methods but does not specify software versions for any libraries, programming languages, or other dependencies.
Experiment Setup Yes For the data degradation process, we set the accuracy levels T to 1000 to create more diverse data, and this data is divided into K = 5 groups for training K models (Section 3.2). We train the KITTI model for 80 epochs and the 3DMatch/3DLo Match models for 20 epochs, taking around 48 hours on a single V100 GPU. Within each refinement network, the attention module is applied three times (i.e., L = 3 in Fig. 2).