Riemannian Local Mechanism for SPD Neural Networks
Authors: Ziheng Chen, Tianyang Xu, Xiao-Jun Wu, Rui Wang, Zhiwu Huang, Josef Kittler
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments involving multiple visual tasks validate the effectiveness of our approach. We evaluate the proposed MSNet in three challenging visual classification tasks |
| Researcher Affiliation | Academia | 1School of Artificial Intelligence and Computer Science, Jiangnan University 2School of Computing and Information Systems, Singapore Management University 3Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey |
| Pseudocode | No | The paper includes Figure 1 which is an "Illustration of the proposed Multi-scale Submanifold Network (MSNet)", but this is a conceptual diagram, not pseudocode or an algorithm block. No explicit "Algorithm" or "Pseudocode" section found. |
| Open Source Code | Yes | The supplement and source code can be found in https://github.com/Git ZH-Chen/MSNet.git. |
| Open Datasets | Yes | Cambridge-Gesture (CG) (Kim and Cipolla 2008) and the UCF-101 (Soomro, Zamir, and Shah 2012) datasets, and skeleton-based action recognition with the First-Person Hand Action (FPHA) (Garcia-Hernando et al. 2018) dataset |
| Dataset Splits | Yes | For this dataset, following the criteria in (Chen et al. 2020), we randomly select 20 and 80 clips for training and testing per class, respectively. For a fair comparison, we follow the protocols in (Wang, Wu, and Kittler 2021). In detail, we use 600 action clips for training and 575 for testing. The seventy-thirty-ratio (STR) protocol is exploited to build the gallery and probes. |
| Hardware Specification | Yes | For training our MSNet, we use an i5-9400 (2.90GHz) CPU with 8GB RAM. |
| Software Dependencies | No | No specific software versions (e.g., Python 3.x, PyTorch 1.x) are mentioned. Only "source code" is mentioned, but not the specific environment dependencies with versions. |
| Experiment Setup | Yes | The initial learning rate is λ = 1e 2 and reduced by 0.8 every 50 epochs to a minimum of 1e 3. Besides, the batch size is set to 30, and the weights in Bi Map layers are initialized as random semi-orthogonal matrices. |