Domain Generalizable Person Search Using Unreal Dataset

Authors: Minyoung Oh, Duhyun Kim, Jae-Young Sim

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that the proposed method provides the competitive performance to existing person search methods even though it is applicable to arbitrary unseen datasets without any prior knowledge and re-training burdens. We used Py Torch for all experiments with a single NVIDIA RTX-3090 GPU. We used the unreal dataset of JTA* only for training, and used the real datasets of CUHK-SYSU (Xiao et al. 2017) and PRW (Zheng et al. 2017) for testing.
Researcher Affiliation Academia Graduate School of Artificial Intelligence, UNIST, Republic of Korea {mmyy2513, duhyunkim, jysim}@unist.ac.kr
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not explicitly provide a link to open-source code or state that the code will be released.
Open Datasets Yes We used the unreal dataset of JTA (Fabbri et al. 2018) obtained from the photo-realistic video game Grand Theft Auto V, where the details for each person instance are automatically annotated such as the bounding boxes, identities, and keypoints. We used the unreal dataset of JTA* only for training, and used the real datasets of CUHK-SYSU (Xiao et al. 2017) and PRW (Zheng et al. 2017) for testing.
Dataset Splits No We constructed the JTA* dataset based on JTA for the purpose of person search by taking 256 sequences in the training category of JTA, which are then divided into 226 sequences for training and 30 sequences for test, respectively. The paper describes training and testing splits but does not explicitly mention a separate validation split for hyperparameter tuning.
Hardware Specification Yes We used Py Torch for all experiments with a single NVIDIA RTX-3090 GPU.
Software Dependencies No We used Py Torch for all experiments with a single NVIDIA RTX-3090 GPU.
Experiment Setup Yes We set the batch size to 4 and used the SGD optimizer with a momentum of 0.9. The warm-up learning rate scheduler linearly increases the learning rate from 0 to 0.003 during the first epoch, and the learning rate decays by multiplying 0.1 every third epoch. We empirically set the weights of losses to 10 and 0.1 for Lfid and Ldom, respectively, and 1 otherwise. τfid = 200 in (1), τid = 1/30 in (4), τdom = 1 in (8), wid = 2/3 in (3), and wdom = 2/3 in (7). During the training phase, we applied the Resize and Horizontal Flip transformations with the probability of 0.5.