Enhance Multi-View Classification Through Multi-Scale Alignment and Expanded Boundary
Authors: Yuena Lin, Yiyuan Wang, Gengyu Lyu, Yongjian Deng, Haichun Cai, Huibin Lin, Haobo Wang, Zhen Yang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiment results across various datasets demonstrate the superiority of the proposed model against existing state-of-the-art methods. [...] We explain the rationality of the model design theoretically, and extensive experimental results across diverse public datasets and comprehensive experimental analysis have verified that the proposed model shows significant superiority against the existing state-of-the-art models. In our experiment, we evaluate our model on eight public multi-view datasets including Hand Written, Scence15, PIE, CCV, Animal, 100Leaves, Hdigit, and Youtube Face. To show the effectiveness of the proposed model, seven state-of-the-art methods are adopted to compare with, including mmdynamics (Han et al. (2022a)), ETMC (Han et al. (2022b)), UMDL (Xu et al. (2023a)), PDMF (Xu et al. (2023b)), IPMVSC (Hu et al. (2023)), MV-HFMD (Black & Souvenir (2024)), and RCML (Xu et al. (2024)). For each dataset, we split 80% instances for training and the remainder for testing. To obtain reliable results, we implement the same dataset split ten times for all the models. More details of the datasets are provided in the Appendix. Table 1 presents the classification results between our model and other baseline models, and we adopt Accuracy (Acc), Purity, Recall, and Macro-F1 as the metrics. [...] 4.2 Ablation Study To evaluate the effectiveness of the multi-scale alignment module and the rationality of the Comb operator determination, we conduct the ablation study whose results are provided in Table 2, and have the following observations: |
| Researcher Affiliation | Collaboration | 1 College of Computer Science, Beijing University of Technology, Beijing 2 Idealism Beijing Technology Co., Ltd., Beijing 3 College of Computer and Data Science, Fuzhou University, Fuzhou 4 School of Software Technology, Zhejiang University, Hangzhou |
| Pseudocode | Yes | Algorithm 1 Training process of MAMC |
| Open Source Code | No | The paper does not provide a direct link to a source-code repository, nor does it contain an explicit statement about the release of the code for the methodology described in this paper. It mentions implementing the model but does not offer access to the code. |
| Open Datasets | Yes | In our experiment, we evaluate our model on eight public multi-view datasets including Hand Written, Scence15, PIE, CCV, Animal, 100Leaves, Hdigit, and Youtube Face. [...] We provide the complete statistics of the used datasets in Table 3. Hand Written 1 and Hdigit 2 datasets are composed of handwritten digit images, PIE 3 and Youtube Face 4 datasets collect the data on facial recognition and expression recognition, Scene15 5 dataset contains images from 15 different scene categories, CCV 6 dataset comprises videos under complex conditions, Animal 7 dataset contains images of different animals, 100Leaves 8 dataset is composed of images from 100 different leaves. 1 Hand Written: http://archive.ics.uci.edu/ml/datasets/Multiple+Features. 2 Hdigit: https://archive.ics.uci.edu/ml/index.php. 3 PIE: https://www.cs.cmu.edu/afs/cs/project/PIE. 4 Youtube Face: https://www.cs.tau.ac.il/wolf/ytfaces. 5 Scene15: http://www-cvr.ai.uiuc.edu/ponce_grp/data/scene15. 6 CCV: http://www.ee.columbia.edu/ln/dvmm/CCV. 7 Animal: https://www.cs.ucf.edu/~xmzhang/datasets. 8 100Leaves: https://archive.ics.uci.edu/ml/datasets/One-hundred+plant+species+leaves+data+set. |
| Dataset Splits | Yes | For each dataset, we split 80% instances for training and the remainder for testing. To obtain reliable results, we implement the same dataset split ten times for all the models. |
| Hardware Specification | Yes | All experiments are conducted on a server with 8 NVIDIA Ge Force 3090 (24 GB memory each). |
| Software Dependencies | No | The proposed model is implemented in Pytorch and trained with an SGD optimizer. The paper mentions software tools like Pytorch and SGD but does not specify their version numbers, which is crucial for reproducibility. |
| Experiment Setup | Yes | The dimension settings of the encoder and decoder are formed by {Dv, 1.4 * 512, 1.2 * 512, 512} and {512, 0.6 * Dv, 0.8 * Dv, Dv} respectively. The number of membership functions L is set as 5. The learning rate is chosen from 1e-3, 3e-3, 5e-3, 1e-2 , and the coefficients α and β are chosen from {0.1, 0.25, 0.5, 0.75, 1}. The temperature coefficients τins and τpro are chosen from {0.5, 0.75, 1.0, 1.5, 2.0} and {0.5, 0.75, 1.0, 1.2, 1.5} respectively. [...] For a fair comparison, we set the same batch size and training epochs for all the models. [...] Algorithm 1 Training process of MAMC Input: Multi-view dataset: D = {({x(v) i }V v=1, yi)|1 i N}; Hyperparameters τins, τpro, α, and β; Training epochs T. Output: Model parameters. |