WaveLoss: An Adaptive Dynamic Loss for Deep Gait Recognition
Authors: Zicheng Wang, Qiuxia Wu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our proposed Waveloss achieves state-of-the-art performance on various gait recognition datasets and can improve the performance of different backbones as well. Evaluation on OU-MVLP. Table 1 shows the experimental results of our Waveloss on the OUMVLP indoor dataset. ... Ablation Study Effect of Hyperparamemter l. To thoroughly investigate the effects of the Dynamic Learning Process, we conducted comparative experiments with different parameter values for l, as shown in Table 4, dividing them into static and dynamic groups. |
| Researcher Affiliation | Academia | Zicheng Wang, Qiuxia Wu* South China University of Technology EMAIL, EMAIL |
| Pseudocode | No | The paper describes methods and formulas (e.g., Formula 4, 5, 6, 7) but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not explicitly state that the source code for their proposed Waveloss is open-source, provide a link to a repository, or mention code in supplementary materials for their own work. |
| Open Datasets | Yes | Dataset To assess the effectiveness of the Waveloss, we will evaluate it on three popular gait datasets, including two indoor datasets, CASIA-B (Yu et al. 2006) and OU-MVLP (Takemura et al. 2018), and one outdoor dataset, GAIT-3D (Zheng et al. 2022b). CASIA-B. The CASIA-B dataset stands as one of the most widely used repositories for gait analysis, comprising data from 124 subjects. ... OUMVLP. The OUMVLP dataset stands as one of the most extensive repositories for gait recognition, boasting a vast collection of data from 10,307 subjects. ... Gait3D. The Gait3D dataset is a substantial repository comprising data from 4,000 subjects |
| Dataset Splits | Yes | Gait3D. ... To ensure compatibility with other algorithms, the dataset is partitioned into a training set containing 3,000 subjects and a test set comprising 1,000 subjects. ... On the CASIA-B dataset, ... Unless otherwise specified, our experimental settings follow the settings of the Open Gait (Fan et al. 2023b). |
| Hardware Specification | Yes | All of our experiments were conducted on the Ge Force RTX 2080 Ti or 2080. |
| Software Dependencies | No | The paper mentions software components like "Adam optimizer" and "Multi Step LR strategy" but does not provide specific version numbers for any libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages. |
| Experiment Setup | Yes | On the CASIA-B dataset, we use Gait GL as the backbone network with channel sizes set to [64, 128, 256] and circle loss as the base loss function. The number of iterations is set to 10,000, with a learning rate of 1e-4. We employ the Multi Step LR strategy, reducing the learning rate by a factor of 10 at iterations 80,000 and 95,000. We utilize the Adam optimizer with a weight decay of 5e-4. Data augmentation techniques such as random flipping and occlusion are also applied. Cosine similarity is used for probe matching. On the OUVMLP dataset, we employ Deep Gait V2 as the backbone network. A triplet loss with a margin of 0.2 and cross-entropy are used as the base loss functions. The number of iterations is 150,000. The learning rate is set to 0.1, with a weight decay of 5e-4. We set the batch size to [32, 8]. Data augmentation strategies include perspective transformations, flipping, and rotation. Probe matching is performed using Euclidean distance. On the Gait3D dataset, we utilize Deep Gait V2 as the backbone network. We employ a triplet loss with a margin of 0.2 and cross-entropy as the base loss functions. The learning rate is set to 0.1, with a weight decay of 5e-4. The total number of iterations is 120,000, and we adopt the Multi Step LR strategy, where the learning rate is reduced by a factor of 1/10 at iterations 40,000, 80,000, and 100,000. Data augmentation strategies include perspective transformations, flipping, and rotation. Probe matching is performed using Euclidean distance. |