Rethinking Pseudo Labels for Semi-supervised Object Detection
Authors: Hengduo Li, Zuxuan Wu, Abhinav Shrivastava, Larry S. Davis1314-1322
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method improves state-of-the-art SSOD performance by 1-2% AP on COCO and PASCAL VOC while being orthogonal and complementary to most existing methods. |
| Researcher Affiliation | Academia | 1 University of Maryland, College Park 2 Fudan University {hdli,abhinav,lsd}@cs.umd.edu, zxwu@fudan.edu.cn |
| Pseudocode | No | No structured pseudocode or clearly labeled algorithm blocks were found. |
| Open Source Code | No | For more details of the implementation such as choices of hyper parameters and training recipes, we refer readers to the ar Xiv version of our paper (arxiv.org/abs/2106.00168). This link refers to an arXiv paper and does not explicitly state the release of source code or provide a direct link to a code repository. |
| Open Datasets | Yes | We evaluate our method on two standard object detection datasets, COCO (Lin et al. 2014) and PASCAL VOC (Everingham et al. 2010), under semi-supervised settings following (Jeong et al. 2019; Sohn et al. 2020b; Liu et al. 2021; Zhou et al. 2021; Yang et al. 2021). |
| Dataset Splits | Yes | In particular, four settings are used: (1) COCO-full: the COCO train2017 set containing 118k images is used as the labeled set, and the additional 123k unlabeled images are used as unlabeled set; (2) COCO-partial: we follow (Sohn et al. 2020b) and randomly sample 1%/2%/5%/10% images from COCO train2017 set as the labeled set, and use the remaining images in train2017 as the unlabeled set; (3) PASCAL VOC: the VOC07 trainval set is used as labeled set and the VOC12 trainval is used as unlabeled set; (4) PASCAL VOC + COCO-20: following (Sohn et al. 2020b), images from COCO containing the 20 classes in PASCAL VOC are used as an additional unlabeled set. For evaluation, the val2017 set of COCO and the VOC07 test set of PASCAL VOC are used. |
| Hardware Specification | No | No specific hardware details (GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running experiments were found. |
| Software Dependencies | No | Our implementation follows existing approaches for fair comparison, and thus we use Faster-RCNN with FPN (Lin et al. 2017) as our detector using a Res Net-50 (He et al. 2016) as its backbone network. This describes models/frameworks but does not provide specific software names with version numbers (e.g., PyTorch 1.9, TensorFlow 2.x). |
| Experiment Setup | No | For more details of the implementation such as choices of hyper parameters and training recipes, we refer readers to the ar Xiv version of our paper (arxiv.org/abs/2106.00168). While some hyperparameter sensitivity is discussed in the ablation study, the full, specific experimental setup details are explicitly deferred to an external resource, indicating they are not fully present in the main text. |