Low-Confidence Samples Mining for Semi-supervised Object Detection

Authors: Guandu Liu, Fangyuan Zhang, Tianxiang Pan, Jun-Hai Yong, Bin Wang

IJCAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On the MS-COCO benchmark, our method achieves 3.54% m AP improvement over state-of-the-art methods under 5% labeling ratios. ... Specifically, LSM introduces an additional branch called pseudo information mining (PIM) for self-learning low-confidence pseudo-labels. ... We conduct a cross-domain task and introduce DDETR baseline into SSOD. ... We carry out extensive experiments to validate the effectiveness of LSM on the MS-COCO [Lin et al., 2014], PASCAL VOC [Everingham et al., 2010], and Image Net [Deng et al., 2009] benchmarks.
Researcher Affiliation Academia Guandu Liu1,2 , Fangyuan Zhang1,2 , Tianxiang Pan1,2 , Jun-Hai Yong1,2 and Bin Wang1,2 1School of Software, Tsinghua University, China 2Beijing National Research Center for Information Science and Technology (BNRist), China {liugd21, zhangfy19}@mails.tsinghua.edu.cn, ptx9363@gmail.com, {yongjh, wangbins}@tsinghua.edu.cn
Pseudocode No The paper does not contain pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide any specific links to open-source code or an explicit statement that the code is publicly available.
Open Datasets Yes In this section, we carry out extensive experiments to validate the effectiveness of LSM on the MS-COCO [Lin et al., 2014], PASCAL VOC [Everingham et al., 2010], and Image Net [Deng et al., 2009] benchmarks.
Dataset Splits Yes MS-COCO contains two training sets, the train2017 dataset with 118K labeled images and the unlabeled2017 dataset with 123K unlabeled images. ... We evaluate the model on COCO-val2017 for (1)(2) and VOC07-test for (3). ... For COCO-standard, the entire training steps are 180, 000, of which the first 20, 000 steps are used to pre-train the student model with labeled images.
Hardware Specification No The paper does not explicitly describe the hardware specifications (e.g., specific GPU or CPU models) used for running its experiments.
Software Dependencies No The paper mentions using 'Faster-RCNN as our base object detector' and 'Deformable-DETR (DDETR)', but it does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For the main branch, we set pseudo boxes filtering threshold t to 0.7. While for LSM, which can have a higher tolerance for pseudo boxes, we set the threshold α to 0.5. ... For COCO-standard, the entire training steps are 180, 000, of which the first 20, 000 steps are used to pre-train the student model with labeled images. ... strong data augmentation involves random jittering, gaussian noise, crop, and weak data augmentation involves random resize and flip.