AdvectiveNet: An Eulerian-Lagrangian Fluidic Reservoir for Point Cloud Processing
Authors: Xingzhe He, Helen Lu Cao, Bo Zhu
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted three parts of experiments, including the ablation tests and the applications for classification and segmentation. We implemented the system in Py Torch (see the submitted source code) and conducted all the tests on a single RTX 2080 Ti GPU. In the ablation tests, we evaluated the functions of the advection module, temporal resolution, grid resolution, and the functions of the PIC/FLIP scheme on Model Net10 (Z. Wu, 2015) and Shape Net (Yi et al., 2016). For classification, we tested our network on Model Net40 and its subset Model Net10. We used the class prediction accuracy as our metric. For segmentation, we tested our network on Shape Net (Yi et al., 2016) and S3DIS data set (Armeni et al., 2016). We used mean Intersection over Union (m Io U) to evaluate our method and compare with other benchmarks. |
| Researcher Affiliation | Academia | Xingzhe He Dartmouth College Rutgers University Helen Lu Cao Dartmouth College Bo Zhu Dartmouth College |
| Pseudocode | No | The paper includes 'Workflow overview' (Figure 2) and 'Network architectures' (Figure 3) which illustrate the process, but it does not provide structured pseudocode or clearly labeled algorithm blocks in a text format. |
| Open Source Code | Yes | We implemented the system in Py Torch (see the submitted source code) and conducted all the tests on a single RTX 2080 Ti GPU. |
| Open Datasets | Yes | In the ablation tests, we evaluated the functions of the advection module, temporal resolution, grid resolution, and the functions of the PIC/FLIP scheme on Model Net10 (Z. Wu, 2015) and Shape Net (Yi et al., 2016)... For classification, we tested our network on Model Net40 and its subset Model Net10... For segmentation, we tested our network on Shape Net (Yi et al., 2016) and S3DIS data set (Armeni et al., 2016). |
| Dataset Splits | Yes | In this part, we further discuss our algorithm and its performance on the large-scale S3DIS dataset (Armeni et al., 2017). Unlike Model Net and Shape Net, the S3DIS consists of colored point clouds collected from real world. We train on the area 1,2,3,4,6 and test on the area 5. |
| Hardware Specification | Yes | We implemented the system in Py Torch (see the submitted source code) and conducted all the tests on a single RTX 2080 Ti GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch' as the implementation framework but does not provide a specific version number. It also cites several algorithms/techniques but without specific software package versions. |
| Experiment Setup | Yes | We use dropout ratio 0.3 on the last fully connected layer before class score prediction. The decay rate for batch normalization starts with 0.5 and is gradually decreased to 0.01. We use adamw optimizer (Loshchilov & Hutter, 2017) with initial learning rate 0.001, weight decay rate 0.005, momentum 0.9 and batch size 32. The learning rate is multiplied by 0.8 every 20 epochs. We train the model for 200 epochs. We use the label smoothing techique (Pereyra et al., 2017) with confidence 0.8. We use the grid size 16 and 32 for classification and segmentation, respectively. |