Diverse Knowledge Distillation for End-to-End Person Search
Authors: Xinyu Zhang, Xinlong Wang, Jia-Wang Bian, Chunhua Shen, Mingyu You3412-3420
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on the CUHKSYSU and PRW datasets demonstrate the superiority of our method against existing approaches it achieves on par accuracy with state-of-the-art two-step methods while maintaining high efficiency due to the single joint model. |
| Researcher Affiliation | Academia | Xinyu Zhang1,2, Xinlong Wang2, Jia-Wang Bian2, Chunhua Shen2,3, Mingyu You1 1Tongji University, China 2The University of Adelaide, Australia 3Monash University, Australia |
| Pseudocode | No | The paper provides architectural diagrams (Figure 1, Figure 2, Figure 3) and describes the methods in text, but it does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at: https://git.io/DKD-Person Search |
| Open Datasets | Yes | We carry out experiments on CUHK-SYSU and PRW datasets. CUHK-SYSU (Xiao et al. 2017) is a large-scale dataset extracted from the street and movie snapshots. ... PRW dataset (Zheng et al. 2017) is collected by six cameras at different locations on a university campus. |
| Dataset Splits | No | The paper describes training and test set splits for CUHK-SYSU and PRW datasets (e.g., "The training set contains 11,206 images... while the test set includes 6,978 gallery images and 2,900 query persons."), but it does not explicitly provide details about a validation dataset split or how it was used to reproduce experiments. |
| Hardware Specification | Yes | All experiments are conducted on a Ge Force GTX 1080 Ti machine. |
| Software Dependencies | No | The paper mentions using "SGD optimizer", "Faster R-CNN", "ResNet-50", and "ImageNet", but it does not provide specific version numbers for any software dependencies like programming languages, libraries (e.g., PyTorch, TensorFlow), or CUDA. |
| Experiment Setup | Yes | We train all models using SGD optimizer with a momentum of 0.9 and a weight decay of 5 10 4. Input images are resized to have at least 800 pixels on the short size and at most 1333 pixels on the long side. We set the batch size to 4. The initial learning rate is 0.001 and then multiplied by 0.1 after 3 104 iterations. The total iteration is 5 105. We set λ = 0.1, βp = 0.1βpr = βtr = 1.0 and m = 0.3 in Eq. (4) and S to 16 in FSA. For the Re-ID model, we use the softmax cross-entropy loss and the SGD optimizer. We resize all the person patches to H W = 256 128. The batch size is set to 64. The initial learning rate is 3.5 10 4 and the total training epoch is 240. Besides, we set p to 10 and α to 40 in the process of ISA. |