Exploring Visual Context for Weakly Supervised Person Search
Authors: Yichao Yan, Jinpeng Li, Shengcai Liao, Jie Qin, Bingbing Ni, Ke Lu, Xiaokang Yang3027-3035
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed framework on the following two datasets. CUHK-SYSU (Xiao et al. 2017) is one of the largest public datasets designed for person search, which contains 18,184 images captured from streets and TV/movie frames. This dataset also includes 96,143 bounding box annotations, with 8,432 different identities. PRW (Zheng et al. 2017) was collected from six surveillance cameras. It contains 11,816 video frames, with 43,110 annotated bounding boxes and 932 identities. Evaluation Protocol. We employ the standard train/test splits for both CUHK-SYSU and PRW. ... We report the mean average precision (m AP) and top-1 ranking accuracy as evaluation metrics. Analytical Results. Comparative Results. We first evaluate the effectiveness of the proposed context-guided learning strategies. We compare the baseline method with different combinations of context information, and report the results on CUHK-SYSU in Table 1. |
| Researcher Affiliation | Collaboration | 1 Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, Shanghai, China. 2 Inception Institute of Artificial Intelligence, Abu Dhabi, United Arab Emirates. 3 University of Chinese Academy of Sciences, Beijing, China. |
| Pseudocode | Yes | Algorithm 1: Clustering with Scene Context |
| Open Source Code | Yes | Our code is available at https://github. com/ljpadam/CGPS |
| Open Datasets | Yes | We evaluate the proposed framework on the following two datasets. CUHK-SYSU (Xiao et al. 2017) is one of the largest public datasets designed for person search... PRW (Zheng et al. 2017) was collected from six surveillance cameras. ... Furthermore, the improvement on PRW is more significant, maybe due to the fact that PRW contains fewer training samples. However, adding more training data does not always bring improvement, e.g., the combination of CUHK-SYSU and COCO (Lin et al. 2014) achieves inferior performance compared with employing CUHK-SYSU alone. |
| Dataset Splits | No | The paper mentions 'standard train/test splits' and provides counts for the training set, but does not specify a separate 'validation split' with percentages or counts, or refer to a standard validation split for the datasets used. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments, only mentioning software like Mind Spore. |
| Software Dependencies | Yes | This project is supported by Mind Spore 1. ... We employ stochastic gradient descent (SGD) optimizer, with the weight decay set to 0.0005. ... We employ DBSCAN (Ester et al. 1996) with self-paced training (Ge et al. 2020) as the clustering method. We set ϵ = 0.7, while other hyperparameters follow SPCL (Ge et al. 2020). |
| Experiment Setup | Yes | Following (Yan et al. 2021), we employ a multiscale training strategy, while input images are resized to a fixed size of 1500 900 for inference. In the training phase, random flipping is applied and we employ stochastic gradient descent (SGD) optimizer, with the weight decay set to 0.0005. We set the batch size to 4 and initialize the learning rate to 0.0012, which is reduced by a factor of 10 at epoch 16, training to a total of 22 epochs. We set the default hyperparameters γ = 0.2, α1 = 1, α2 = 0.25, and λ = 0.6. We employ DBSCAN (Ester et al. 1996) with self-paced training (Ge et al. 2020) as the clustering method. We set ϵ = 0.7, while other hyperparameters follow SPCL (Ge et al. 2020). |