Hierarchical Discriminative Learning for Visible Thermal Person Re-Identification
Authors: Mang Ye, Xiangyuan Lan, Jiawei Li, Pong Yuen
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments illustrate the effectiveness and robustness of the proposed method. |
| Researcher Affiliation | Academia | Mang Ye, Xiangyuan Lan, Jiawei Li, Pong C. Yuen Department of Computer Science, Hong Kong Baptist University {mangye,jwli,pcyuen}@comp.hkbu.edu.hk, xiangyuanlan@life.hkbu.edu.hk |
| Pseudocode | No | The paper describes algorithmic steps but does not include a clearly labeled pseudocode block or algorithm figure. |
| Open Source Code | Yes | Code is available on the first author s website. |
| Open Datasets | Yes | We use the publicly available Reg DB dataset (Nguyen et al. 2017) for evaluation. |
| Dataset Splits | Yes | All the experiments are conducted following the standard evaluation protocol in existing visible image based re-id works, i.e., we randomly split the datasets into two halves, one for training and the other for testing. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models or processor types used for running the experiments. |
| Software Dependencies | No | The paper mentions the use of 'Tensorflow' but does not specify its version or the versions of other software dependencies. |
| Experiment Setup | Yes | For two-stream CNN feature learning, we implement it on Tensorflow, we set the length of fc2 as 2048, which could achieve a slightly better performance than the original setting (4046). The trade-off parameter α of identity loss and contrastive loss it set as 0.2, and the maximum number of training epochs is set to 30. For efficiency considerations, we firstly reduce the dimension of features generated by fc2 from 2048 to 600 by PCA. For HCML, we set Nd as 600 to keep all the energies. Empirically, the balancing parameter β of modality-specific and modality-shared term is set to 0.2, since the major difficulty for cross-modal matching problem is to handle the cross-modality discrepancy. The smooth parameter γ of the logistic loss function is set to 1. And the learning rate to update V and T is initialized l0 = 0.1, k is set to 0.9 in all our experiments. |