Multi-View Domain Adaptive Object Detection on Camera Networks
Authors: Yan Lu, Zhun Zhong, Yuanchao Shu
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on two real-world multi-camera datasets demonstrate significant advantages of our approach over the state-of-the-art domain adaptive object detection methods. |
| Researcher Affiliation | Academia | Yan Lu1, Zhun Zhong2, Yuanchao Shu3 1Department of Electrical Engineering and Computer Sciences, New York University 2Department of Information Engineering and Computer Science, University of Trento 3College of Control Science and Engineering, Zhejiang University |
| Pseudocode | No | The paper describes its methods textually and with diagrams but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | A complete version with a technical appendix is available on the website: https://jason-cs18.github.io. |
| Open Datasets | Yes | In this paper, we conduct experiments on one general dataset (MS-COCO (Lin et al. 2014)) and two real-world multi-camera datasets (Wild Track (Chavdarova et al. 2018) and City Flow (Tang et al. 2019)). |
| Dataset Splits | Yes | For evaluation, we set the ratio of the training set, evaluation set, and testing set to 16 : 4 : 5 for both Wild Track and City Flow. |
| Hardware Specification | No | The paper describes the models used (YOLOv3, Faster R-CNN) and the toolbox (mmdetection) but does not specify the hardware (e.g., GPU, CPU models) used for experiments. |
| Software Dependencies | No | The paper mentions software like 'mmdetection (Chen et al. 2019) toolbox', 'Siam Mask-E (Xin and K 2019)', 'SBS (He et al. 2020)', and 'Vehicle Net (Zheng et al. 2020b)'. However, it does not provide specific version numbers for these software components, which are necessary for reproducible descriptions. |
| Experiment Setup | Yes | In pseudo-label construction, we set Tthr to 0.7. During training, we choose Adam (Kingma and Ba 2015) as the optimizer and set the learning rate to 0.01. The batch size is set to 8. The threshold of classification score (cthr) is set to 0.5. We train the model with 60 epochs in total, in which self-supervised multiview training and single-view detection head fine-tuning are trained with 30 epochs individually. |