SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network
Authors: Ziming Nie, Qiao Wu, Chenlei Lv, Siwen Quan, Zhaoshuai Qi, Muze Wang, Jiaqi Yang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the superior performance of our method through both quantitative and qualitative experiments, showing overall superiority against both existing selfsupervised and supervised methods. |
| Researcher Affiliation | Academia | 1Northwestern Polytechnical University, Xi an, China 2Shenzhen University, Shenzhen, China 3Chang an University, Xi an, China EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the proposed method in detail with figures and text, but it does not contain a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | Code https://github.com/hapifuzi/spu-imr |
| Open Datasets | Yes | We employ the dataset (Zhao et al. 2022, 2023) used for training and testing. It consists of 4719 point clouds split into 3672 training samples and 1047 testing samples. Each sample contains 2048 points, and we further down-sample to 1024 points through FPS to form the test set and the training set. It is noteworthy that the test samples (Zhao et al. 2022, 2023) used are from the dataset adopted by (Yu et al. 2018b; Ye et al. 2021). [...] Results on PU1K under Different Upsampling Ratios. PU1K dataset (Qian et al. 2021a) consists of 1147 samples. [...] Results on the Real-world KITTI Dataset. The KITTI dataset (Geiger et al. 2013) is a benchmark for autonomous driving and computer vision. |
| Dataset Splits | Yes | It consists of 4719 point clouds split into 3672 training samples and 1047 testing samples. |
| Hardware Specification | Yes | The training process is conducted on a server with four Ge Force RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions using an "Adam W optimizer (Loshchilov and Hutter 2017)" and "cosine learning rate decay (Loshchilov and Hutter 2016)", which are algorithms or techniques. However, it does not provide specific software versions for programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | During the training phase, each sparse point cloud (N = 1024) is partitioned into 256 patches, and each patch contains 16 points. The mask ratio m is assigned to 0.6. The number of iteration layers L is set to 3. As for the testing phase, we assign q = 8 mask sequences. These mask sequences are generated randomly. [...] The networks are trained for 6000 epochs with a batch size of 128. For training details, we use an Adam W optimizer (Loshchilov and Hutter 2017) and cosine learning rate decay (Loshchilov and Hutter 2016). The initial learning rate is set to 0.001, with a weight decay of 0.05. |