Learning Incremental Triplet Margin for Person Re-Identification

Authors: Yingying Zhang, Qiaoyong Zhong, Liang Ma, Di Xie, Shiliang Pu9243-9250

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on Market-1501, CUHK03, and Duke MTMCre ID show that our approach yields a performance boost and outperforms most existing state-of-the-art methods.
Researcher Affiliation Industry Yingying Zhang, Qiaoyong Zhong, Liang Ma, Di Xie, Shiliang Pu Hikvision Research Institute {zhangyingying7,zhongqiaoyong,maliang6,xiedi,pushiliang}@hikvision.com
Pseudocode Yes Algorithm 1 Global Hard Identity Searching.
Open Source Code No The paper does not provide a specific link or explicit statement about releasing the source code for the described methodology.
Open Datasets Yes We evaluate the proposed approach on three large-scale person Re ID datasets, namely Market-1501 (Zheng et al. 2015), CUHK03 (Li et al. 2014) and Duke MTMC-re ID (Ristani et al. 2016; Zheng, Zheng, and Yang 2017).
Dataset Splits No The paper specifies training and testing splits for datasets (e.g., '12,936 images from 751 identities (including 1 background category) for training and 19,732 images from 750 identities for testing' for Market-1501), but does not explicitly mention or detail a separate validation set split.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper mentions 'Py Torch' and 'Adam optimizer' but does not provide specific version numbers for these or any other software dependencies, which is required for reproducibility.
Experiment Setup Yes The training images are randomly cropped with a ratio uniformly sampled from [0.8, 1) and resized to 288 × 144. Random erasing (Zhong et al. 2017b) and random flipping are applied on resized images with a probability of 0.5. The hyper-parameters of random erasing data augmentation are set the same as (Zhong et al. 2017b). The number of persons P per-batch and number of images per-person K are set to 20 and 4 respectively. Hence, the mini-batch size is 80. For LITM, the base and incremental margins are set as m0 = 4, m1 = 7, m2 = 10. We use the Adam optimizer (Kingma and Ba 2014) with ϵ = 10 3, β1 = 0.99 and β2 = 0.999. The network is trained for 300 epochs in total. And a piecewise learning rate schedule is utilized, where it is fixed to 2 × 10 4 in the first 150 epochs and decayed exponentially in the rest 150 epochs.