Unbiased Teacher for Semi-Supervised Object Detection

Authors: Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, Peter Vajda

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We benchmark Unbiased Teacher with SSL setting using the MS-COCO and PASCAL VOC datasets, namely COCO-standard, COCO-additional, and VOC. When using only 1% labeled data from MS-COCO (COCO-standard), Unbiased Teacher achieves 6.8 absolute m AP improvement against the state-of-the-art method, STAC (Sohn et al., 2020b). Unbiased Teacher consistently achieves around 10 m AP improvements when using only 0.5, 1, 2, 5% of labeled data on MS-COCO. We also provide an ablation study to verify the effectiveness of each proposed component.
Researcher Affiliation Collaboration Yen-Cheng Liu1,2 , Chih-Yao Ma2, Zijian He2, Chia-Wen Kuo1, Kan Chen2, Peizhao Zhang2, Bichen Wu2, Zsolt Kira1, Peter Vajda2 1Georgia Tech, 2Facebook Inc. {ycliu,cwkuo,zkira}@gatech.edu, {cyma,zijian,kanchen18,stzpz,wbc,vajdap}@fb.com
Pseudocode No The paper describes methods and processes but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes 1Code: https://github.com/facebookresearch/unbiased-teacher.
Open Datasets Yes Datasets. We benchmark our proposed method on experimental settings using MS-COCO (Lin et al., 2014) and PASCAL VOC (Everingham et al., 2010) following existing works (Jeong et al., 2019; Sohn et al., 2020b).
Dataset Splits Yes Specifically, there are three experimental settings: (1) COCO-standard: we randomly sample 0.5, 1, 2, 5, and 10% of labeled training data as a labeled set and use the rest of the data as the training unlabeled set. Figure 2: Validation Losses of our model and the model trained with labeled data only. Figure 9: (a) Validation AP and (b) number of pseudo-label bounding boxes per image with various pseudo-labeling thresholds δ.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for running the experiments.
Software Dependencies No The paper states 'Our implementation builds upon the Detectron2 framework (Wu et al., 2019)' but does not specify version numbers for Detectron2 or any other software dependencies.
Experiment Setup Yes Implementation Details. For a fair comparison, we follow STAC (Sohn et al., 2020b) to use Faster RCNN with FPN (Lin et al., 2017a) and Res Net-50 backbone (He et al., 2016) as our object detectior, where the feature weights are initialized by the Image Net-pretrained model, same as existing works (Jeong et al., 2019; Sohn et al., 2020b). We use confidence threshold δ = 0.7. ... Hyper-parameters. We use confidence threshold δ = 0.7 to generate pseudo-labels for all our experiments, the unsupervised loss weight λu = 4 is applied for COCO-standard and VOC, and the unsupervised loss weight λu = 2 is applied for COCO-additional. We apply α = 0.9996 as the EMA rate for all our experiments. Hyper-parameters used are summarized in Table 5. Training. ... We use the SGD optimizer with a momentum rate 0.9 and a learning rate 0.01, and we use constant learning rate scheduler. The batch size of supervised and unsupervised data are both 32 images. For the COCO-standard, we train 180k iterations...