GeoNLF: Geometry guided Pose-Free Neural LiDAR Fields
Authors: Weiyi Xue, Zehan Zheng, Fan Lu, Haiyun Wei, Guang Chen, changjun jiang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on Nu Scenes and KITTI-360 datasets demonstrate the superiority of Geo NLF in both novel view synthesis and multi-view registration of low-frequency large-scale point clouds. |
| Researcher Affiliation | Academia | Weiyi Xue Tongji University xwy@tongji.edu.cn Zehan Zheng Tongji University zhengzehan@tongji.edu.cn Fan Lu Tongji University lufan@tongji.edu.cn Haiyun Wei Tongji University 2311399@tongji.edu.cn Guang Chen Tongji University guangchen@tongji.edu.cn Changjun Jiang Tongji University cjjiang@tongji.edu.cn |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is availiable at https://github.com/ispc-lab/Geo NLF. |
| Open Datasets | Yes | We conducted experiments on two public autonomous driving datasets: Nu Scenes [9] and KITTI-360 [30] dataset, each with five representative Li DAR point cloud sequences. |
| Dataset Splits | Yes | We selected 36 consecutive frames at 2Hz from keyframes as a single scene for Nu Scenes, holding out 4 samples at 9-frame intervals for NVS evaluation. KITTI-360 has an acquisition frequency of 10Hz. We used 24 consecutive frames sampled every 5th frame to match scene sizes of Nuscenes, holding out 3 samples at 8-frame intervals for evaluation. |
| Hardware Specification | No | The paper mentions that the optimization is implemented on PyTorch, but it does not provide specific hardware details such as GPU or CPU models, or memory specifications used for running the experiments. |
| Software Dependencies | No | The optimization of Geo NLF is implemented on Pytorch [42] with Adam [26] optimizer. No specific version numbers for PyTorch or other software dependencies are provided. |
| Experiment Setup | Yes | The entire point cloud scene is scaled within the unit cube space. The optimization of Geo NLF is implemented on Pytorch [42] with Adam [26] optimizer. All the sequences are trained for 60K iterations. Our Geometry optimizer s lr for translation and rotation is the same as the lr for pose in Ne RF with synchronized decay. We use the coarse-to-fine strategy[31, 21], which starts from training progress 0.1 to 0.8. The reweight coefficient for the top-5 frames linearly increases from 0.15 to 1 during training. After every m1 epoch of bundle adjusting global optimization, we proceed with m2 epoch of pure geometric optimization, where m2/m1 decrease from 10 to 1. |