3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration
Authors: Liyuan Zhang, Le Hui, qi liu, Bo Li, Yuchao Dai
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on two public benchmarks, Scan2CAD and ROBI, show that our method achieves a new state-of-the-art performance on the multi-instance point cloud registration task. |
| Researcher Affiliation | Academia | Liyuan Zhang, Le Hui , Qi Liu, Bo Li, Yuchao Dai School of Electronics and Information, Northwestern Polytechnical University Shaanxi Key Laboratory of Information Acquisition and Processing zhangliyuannpu@mail.nwpu.edu.cn, {huile, liuqi, libo, daiyuchao}@nwpu.edu.cn |
| Pseudocode | No | The paper describes the proposed method in textual form but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The project page is at https://npucvr.github.io/3DFMNet/. (Also: "We will release the codes." from NeurIPS checklist.) |
| Open Datasets | Yes | For multi-instance point cloud registration, We train and evaluate our method on two public benchmarks: Scan2CAD [3] and ROBI [43]. |
| Dataset Splits | Yes | For the total of 2,184 pairs of point clouds, it allocates 70% of pairs for training, 10% for validation, and reserves 20% for testing. (Also: There are a total of 4,880 pairs of ROBI, divided into 70% for training, 10% for validation, and 20% for testing.) |
| Hardware Specification | Yes | Our method is trained on NVIDIA RTX 4090 GPUs and uses the Pytorch deep learning platform. |
| Software Dependencies | No | The paper mentions using "Pytorch deep learning platform" and other tools like "KPConv" and "DBSCAN" but does not provide specific version numbers for any of the software dependencies. |
| Experiment Setup | Yes | We employ the Adam optimizer for 60 epochs. In initial learning rate and weight decay are set to 0.001 and 0.0001, respectively. We use a KPConv-FPN[37] backbone followed by [46] for feature extraction. We utilize a voxel subsampling approach to reduce the resolution of the point clouds, resulting in the creation of sampled points and dense points, which are then inputted into the network. The initial step involves downsampling the input point clouds using a voxel-grid filter with a size of 2.5cm for Scan2CAD and 0.15cm for ROBI. Subsequently, we employ a 4-stage backbone architecture in both the multi-object focusing and sub-matching network. ... The search radius is set to 1.2 times the size of the CAD model. Specifically, we randomly sample 4096 points from the dense points obtained through voxel subsampling in both Scan2CAD and ROBI datasets. |