Uncertainty Modeling with Second-Order Transformer for Group Re-identification
Authors: Quan Zhang, Jian-Huang Lai, Zhanxiang Feng, Xiaohua Xie3318-3325
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | A large number of experiments have been conducted on three available datasets, including CSG, Duke Group and Road Group, which show that the proposed SOT outperforms all previous state-of-the-art methods. |
| Researcher Affiliation | Academia | Quan Zhang1, Jian-Huang Lai1,2,3,4*, Zhanxiang Feng1, Xiaohua Xie1,2,3 1School of Computer Science and Engineering, Sun Yat-Sen University, Guangzhou, China 2Guangdong Key Laboratory of Information Security Technology, Guangzhou, China 3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China 4Key Laboratory of Video and Image Intelligent Analysis and Applicaiton Technology, Ministry of Public Security, China |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any concrete access information (e.g., specific repository link, explicit code release statement) for its source code. |
| Open Datasets | Yes | The proposed SOT is evaluated on Duke Group (Lin et al. 2021), Road Group (Lin et al. 2021) and CSG (Yan et al. 2020) datasets. |
| Dataset Splits | No | The paper mentions training and testing splits, for example: 'the training and testing set of Duke Group and Road Group are randomly and equally split' and '859/699 groups are split for training/testing' for CSG, but does not explicitly define a separate validation split. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types) used for running its experiments. |
| Software Dependencies | No | The paper mentions using 'Vi T-Base' and 'SGD' but does not provide specific version numbers for software dependencies or programming languages (e.g., Python, PyTorch). |
| Experiment Setup | Yes | We crop all the member patches by the given bounding box and resize them to 256x128. In the training stage, the random horizontal flip and random erasing are performed with a fixed probability of 0.5. Each mini-batch is sampled with 16 group identities, and each group identity selects 4 images. We choose SGD (Bottou 2012) as the optimizer. Our training stage ends when the iteration number reaches 400 epochs. We use a cosine annealing learning rate strategy. The initial learning rate is 2e-3, and the minimum learning rate is 1.6e-4. The learning rate of the inter-member module is multiplied by 0.1. The weight decay is 1e-4. The selection of hard samples in triplet loss adopts an online mining strategy. In the testing stage, we do not use any data augmentation and re-ranking. We use the Euclidean distance to measure the normalized features. |