Dual-frame Fluid Motion Estimation with Test-time Optimization and Zero-divergence Loss
Authors: Yifei Zhang, Huan-ang Gao, zhou jiang, Hao Zhao
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct comprehensive evaluations using different data domains on our proposed framework. First, we compare our method with SOTA fully supervised methods (Sec. 4.1). Next, we examine its performance under the constrained size of training data, reflecting real-world situations where domain-specific data is limited (Sec. 4.2). We then assess the framework s performance under different domains with increasing domain shift, highlighting its cross-domain robustness (Sec. 4.3). Additionally, we conduct comprehensive ablation studies on the components of our framework (Sec. 4.4) to validate their effects. |
| Researcher Affiliation | Academia | Yifei Zhang University of Chinese Academy of Sciences zhangyifei21a@mails.ucas.ac.cn Huan-ang Gao AIR, Tsinghua University gha24@mails.tsinghua.edu.cn Zhou Jiang Beijing Institute of Technology jzian@bit.edu.cn Hao Zhao AIR, Tsinghua University Beijing Academy of Artificial Intelligence zhaohao@air.tsinghua.edu.cn |
| Pseudocode | No | The paper describes algorithms and methods but does not present them in a pseudocode block or a clearly labeled Algorithm section. |
| Open Source Code | Yes | Code, data and models are available at https://github.com/Forrest-110/Fluid Motion Net. |
| Open Datasets | Yes | Following the previous SOTA method Got Flow3D [52], our datasets include Fluid Flow3D [52] and its six fluid cases, Deformation Flow [98] and AVIC [41]. The Fluid Flow3D [53] is a large synthetic dataset designed for the study of 3D fluid flow. Specifically, it offers enough data for training and serves as a benchmark to evaluate the flow estimation capabilities of supervised 3D fluid flow motion learning techniques. |
| Dataset Splits | No | The paper defines training and test sets but does not explicitly mention a separate validation set split (with percentages or counts) for hyperparameter tuning or model selection during training. |
| Hardware Specification | Yes | Time profiling is conducted on a single RTX 3090 Ti. A default learning rate of 0.001 is set and the training is run on a single RTX 4070TI. |
| Software Dependencies | No | The paper mentions software components like Adam optimizer and network architectures (PointNet++, Geo Set Conv, Edge Conv) but does not provide specific version numbers for these or other key software dependencies (e.g., Python, PyTorch, CUDA versions) needed for replication. |
| Experiment Setup | Yes | For the feature extractor, the K-value of our K-nearest-neighbor is chosen to be 32, the embedding-dim to be 128, and the dropout rate to be 0.5. The grid size for splatting is 10 x 10 x 10. LOSS TERM The selected number of neighboring points for the reconstruction loss Lrecon is 32. Likewise, the number of neighboring points for smooth flow loss Lsmooth is 32, and for zero-divergence loss Ldiv is 2. λconf is 0.1, λsmooth is 10, and λdiv is 0.1. During the training phase, we utilize a mini-batch training process with a batch size of 4. To achieve convergence, we train the full-data model and 10%-data model for 100 epochs, and 1%-data model for 300 epochs. A default learning rate of 0.001 is set and the training is run on a single RTX 4070TI. During the test phase, DVE runs for 150 steps with an update rate of 0.01. |