Multi-Prompts Learning with Cross-Modal Alignment for Attribute-Based Person Re-identification
Authors: Yajing Zhai, Yawen Zeng, Zhiyong Huang, Zheng Qin, Xin Jin, Da Cao
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on the existing attribute-involved Re ID datasets, namely, Market1501 and Duke MTMC-re ID, demonstrate the effectiveness and rationality of the proposed MP-Re ID solution. Experiments Experimental Settings Dataset. In this paper, we evaluate the proposed MP-Re ID method on two well-known Re ID datasets: Market1501 (Zheng et al. 2015), Duke MTMC-re ID (Zheng, Zheng, and Yang 2017), as well as the attribute datasets associated with these two datasets, which were manually annotated (Lin et al. 2019). Evaluation Protocols. To evaluate the performance of our approach, we employed Rank@k and mean Average Precision (m AP) as the evaluation metrics for all experiments on the two datasets (Wang et al. 2021; Farooq et al. 2022). Higher values indicate better performance. Implementation Details. We apply our method on a server equipped with the NVIDIA Ge Force RTX 2080 Ti GPU. |
| Researcher Affiliation | Academia | Yajing Zhai1,2*, Yawen Zeng1*, Zhiyong Huang3, Zheng Qin1 , Xin Jin2 , Da Cao1 1College of Computer Science and Electronic Engineering, Hunan University, Changsha, China 2Ningbo Institute of Digital Twin, Eastern Institute of Technology, Ningbo, China 3National University of Singapore, NUS Research Institute in Chongqing {yajingzhai9,yawenzeng11}@gmail.com, huangzy@comp.nus.edu.sg, zqin@hnu.edu.cn, jinxin@eitech.edu.cn, caoda0721@gmail.com |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Meanwhile, we collect a prompt-related Re ID dataset containing multiple attribute prompts about the same person, and we have released the dataset to facilitate the research community1. 1https://github.com/zyj20/MPRe ID. |
| Open Datasets | Yes | In this paper, we evaluate the proposed MP-Re ID method on two well-known Re ID datasets: Market1501 (Zheng et al. 2015), Duke MTMC-re ID (Zheng, Zheng, and Yang 2017), as well as the attribute datasets associated with these two datasets, which were manually annotated (Lin et al. 2019). |
| Dataset Splits | No | The paper uses well-known datasets (Market1501, Duke MTMC-re ID) which have standard splits, but it does not explicitly state the percentages, counts, or how the data was split into training, validation, and testing sets within the paper. |
| Hardware Specification | Yes | We apply our method on a server equipped with the NVIDIA Ge Force RTX 2080 Ti GPU. |
| Software Dependencies | No | The paper mentions using 'Transformer-based models' and 'CLIP' and 'Chat GPT 3.5', but it does not specify software dependencies with version numbers like Python, PyTorch, or CUDA versions. |
| Experiment Setup | Yes | Implementation Details. We use the Transformer-based models and the learning rate is 5 10 7 with a linearly growing. And the warming up is set to 10 to make the model converge faster. In our implementation, we set S = 16 and K = 4 to enable our model to learn from multiple identities and samples per identity. For feature extraction, prompt features and image features are represented as 512-dimensional vectors. Furthermore, we set the ID loss balance factor λid to 0.25 as a regularization strategy, λtri and the weight of Lalign is set to 1. Regarding the triplet loss, we set the margin parameter ξ to 0.3 to create an adequate boundary between the positive and negative samples. |