Temporal-Enhanced Convolutional Network for Person Re-Identification
Authors: Yang Wu, Jie Qiu, Jun Takamatsu, Tsukasa Ogasawara
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on the most widely used benchmark datasets demonstrate the superiority of our proposal, in comparison with the state-of-the-art. |
| Researcher Affiliation | Academia | Yang Wu, Jie Qiu, Jun Takamatsu, Tsukasa Ogasawara Nara Institute of Science and Technology 8916-5 Takayama-cho, Ikoma, Nara, 630-0192, Japan yangwu@rsc.naist.jp, {qiu.jie.qf3, j-taka, ogasawar}@is.naist.jp |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | No | The paper does not provide any explicit statements about open-sourcing code or links to a code repository. |
| Open Datasets | Yes | We evaluate our proposed model (T-CN) on two commonly used benchmark datasets: PRID 2011 (Hirzer et al. 2011) and i LIDS-VID (Wang et al. 2014), for a fair comparison with the state-of-the-art and a detailed empirical study of our proposal. |
| Dataset Splits | No | In our experiments, we randomly generated 10 different splits for each dataset, having 50 percent of people for training and the other 50 percent for testing. |
| Hardware Specification | Yes | RCN takes 0.3941 seconds while T-CN takes 0.3979 seconds with the same Nvidia Titan X GPU ( 1% difference). |
| Software Dependencies | No | The paper does not specify version numbers for any software dependencies. |
| Experiment Setup | Yes | We set the margin m in the verification loss to 2 and the dimensionality of embedded feature space d (the final output of Spatial Conv Net) to 256. The number of convolution kernels k for Temporal Conv Net was set to 32. We trained the whole network for 2,000 epochs with the learning rate set to 1e-6. During Training, we use the same 16 consecutive frames for each sequence as RCN did. During testing, given each probe sequence, we computed its deep features using trained network and used Euclidean distance for the result ranking. Data augmentation (cropping and flipping for each cropped image) was applied to both training and testing as done in the RCN work. |