De-biased Teacher: Rethinking IoU Matching for Semi-supervised Object Detection
Authors: Kuo Wang, Jingyu Zhuang, Guanbin Li, Chaowei Fang, Lechao Cheng, Liang Lin, Fan Zhou
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that the proposed De-biased Teacher consistently outperforms other state-of-the-art methods on the MS-COCO and PASCAL VOC benchmarks. |
| Researcher Affiliation | Academia | 1School of Computer Science and Engineering, Research Institute of Sun Yat-sen University in Shenzhen, Sun Yat-sen University, Guangzhou, China 2School of Artificial Intelligence, Xidian University, Xi an, China 3Zhejiang Lab, Zhejiang, China |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Source codes are available at https://github.com/wkfdb/De-biased-Teracher. |
| Open Datasets | Yes | We evaluate our proposed approach on three object detection datasets, PASCAL VOC (Everingham et al. 2010), MS-COCO (Lin et al. 2014) and Object365 (Shao et al. 2019). |
| Dataset Splits | Yes | Partially labeled COCO: We randomly sample 1%/5%/10% images from train2017 as the labeled dataset and use the remaining images as the unlabeled dataset. ... COCO-val2017 is used as the evaluation set for both 2) and 3). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments, only mentioning that experiments are implemented based on MMDetection. |
| Software Dependencies | No | The paper mentions that experiments are "implemented based on MMDetection (Chen et al. 2019)", but does not provide specific version numbers for MMDetection or any other software dependencies (e.g., Python, PyTorch, CUDA). |
| Experiment Setup | Yes | For fair comparisons, we utilize Faster-RCNN as our detector and use Resnet-50-FPN as the backbone. Following existing works (Xu et al. 2021), we set the EMA updating rate α = 0.999. For the coefficient of unsupervised loss, we set λ = 4 on partially labeled COCO and λ = 2 on fully labeled COCO, VOC and open scene. The similar weak-strong data augmentation schemes in (Liu et al. 2021) are utilized. For RPN and regression loss, the pseudo label is generated by the conventional thresholding method with σRP N = 0.7 and σreg = 0.9. For ROI classification loss, we set threshold δ = 0.05 for selecting foregrounds and distribution refinement. The batch size of labeled and unlabeled data is (12,24) for VOC, (8,32) for partially labeled COCO, (32,32) for fully labeled COCO and open scene. The training iteration is 90k for VOC, 180k for partially labeled COCO and 720k for fully labeled COCO and open scene. |