Unpaired Point Cloud Completion on Real Scans using Adversarial Training
Authors: Xuelin Chen, Baoquan Chen, Niloy J. Mitra
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present quantitative and qualitative experimental results on several noisy and partial datasets. First, we present results on real-world datasets, demonstrating the effectiveness of our method on unpaired raw scans. Second, we thoroughly compare our method to various baseline methods on 3D-EPN dataset, which contains simulated partial scans and corresponding ground truth for full evaluation. Finally, we derive a synthetic noisy-partial scan dataset based on Shape Net, on which we can evaluate the performance degradation of applying supervised methods to test data of different distribution and the performance of our method under varying levels of incompleteness. A set of ablation studies is also included to evaluate our design choices. |
| Researcher Affiliation | Collaboration | Xuelin Chen Shandong University University College London Baoquan Chen Peking University Niloy J. Mitra University College London Adobe Research London |
| Pseudocode | No | The paper describes the network architecture and training process in text and figures, but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an unambiguous statement about releasing source code or a direct link to a code repository for the methodology described. |
| Open Datasets | Yes | Datasets. (A) Real-world dataset comes from three sources. First, a dataset of 550 chairs and 550 tables extracted from the Scan Net dataset split into 90%-10% train-test sets. Second, a dataset of 20 chairs and 20 tables extracted from the Matterport3D dataset. ... Third, a dataset containing cars from the KITTI Velodyne point clouds. (B) 3D-EPN dataset provides simulated partial scans with corresponding ground truth. ... (C) Clean and complete point set dataset contains virtually scanned point sets of Shape Net models covering 8 categories... |
| Dataset Splits | No | For Scan Net, it says 'split into 90%-10% train-test sets.' For synthetic data, 'split them into 90%/10% train/test sets.' It specifies train and test splits but does not explicitly provide a separate validation dataset split percentage or count for reproduction, though 'validation' is mentioned as a general process. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models, memory, or processing units used for running the experiments. |
| Software Dependencies | No | The paper describes network architectures (e.g., MLP, convolutions) and optimizers (Adam) but does not provide specific version numbers for any software frameworks, libraries, or dependencies used in the implementation. |
| Experiment Setup | Yes | For training the AE, we use Adam optimizer with an initial learning rate of 0.0005, β1 = 0.9 and a batch size of 200 and train for a maximum of 2000 epochs. For training the generator and discriminator on the latent spaces, we use Adam optimizer with an initial learning rate of 0.0001, β1 = 0.5 and a batch size of 24 and train the generator and discriminator alternately for a maximum of 1000 epochs. |