Generalizable Person Re-identification via Balancing Alignment and Uniformity

Authors: Yoonki Cho, Jaeyoon Kim, Woo Jae Kim, Junsik Jung, Sung-eui Yoon

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results demonstrate that BAU effectively exploits the advantages of data augmentation, which previous studies could not fully utilize, and achieves state-of-the-art performance without requiring complex training procedures.
Researcher Affiliation Academia Yoonki Cho Jaeyoon Kim Woo Jae Kim Junsik Jung Sung-Eui Yoon
Pseudocode No I could not find any structured pseudocode or algorithm blocks within the paper.
Open Source Code Yes The code is available at https://github.com/yoonkicho/BAU.
Open Datasets Yes We conduct experiments using the following datasets: Market-1501 [95], MSMT17 [83], CUHK02 [43], CUHK03 [44], CUHK-SYSU [85], PRID [31], GRID [55], VIPe R [26], and i LIDs [97], with dataset statistics shown in Table 1.
Dataset Splits No Table 2: Evaluation protocols. Setting Training Data Testing Data Protocol-1 Full-(M+C2+C3+CS) PRID, GRID, VIPe R, i LIDs Protocol-2 M+MS+CS C3 M+CS+C3 MS MS+CS+C3 M Protocol-3 Full-(M+MS+CS) C3 Full-(M+CS+C3) MS Full-(MS+CS+C3) M. While detailed training and testing data splits are provided, an explicit 'validation' split with percentages or counts is not mentioned.
Hardware Specification Yes We implement our framework in Py Torch [64] and utilize two RTX-3090 GPUs for training.
Software Dependencies No We implement our framework in Py Torch [64] and utilize two RTX-3090 GPUs for training. We train the model for 60 epochs using Adam [38] with a weight decay of 5 10 4. (Software names are mentioned but specific version numbers are not provided.)
Experiment Setup Yes Following previous studies [34, 50, 51, 86, 90], we use Res Net-50 [29] pre-trained on Image Net [13] with instance normalization layers as our backbone. All images are resized to 256 × 128. For each iteration, we sample 256 images, consisting of 64 identities with 4 instances for each identity. The total batch size during training is 512, including both original and augmented images. Random flipping, cropping, erasing [101], Rand Augment [11], and color jitter are used for data augmentation. We train the model for 60 epochs using Adam [38] with a weight decay of 5 × 10−4. The initial learning rate is set to 3.5 × 10−4 and is decreased by a factor of 10 at the 30th and 50th epochs. A warmup strategy is applied during the first 10 epochs. The momentum μ is set to 0.1. We empirically set the weighting parameter λ to 1.5 and k for the weighting strategy to 10.