.png){kind=icon}
.png){kind=icon}
Return
DRCO: Dense-Label Refinement and Cross Optimization for Semi-Supervised Object Detection
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2024.3524029.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In semi-supervised object detection (SSOD), the methods based on dense pseudo-labeling bypass complex post-processing while maintaining competitive performance compared to the methods based on sparse pseudo-labeling. However, there are still relatively few researches focused on the dense pseudo-labeling paradigm. In this work, we first experimentally point out the shortcomings of current dense pseudo-labeling methods: 1) Low-quality sampling: the fixed threshold strategies can result in numerous false negatives and false positives. 2) Inconsistency between classification scores and localization quality: classification scores cannot represent localization quality, resulting in poor quality of predicted bounding boxes for sampled positive samples. 3) Suboptimal training approach: current training methods only utilize its knowledge from the perspective of final dense pseudo-labels, failing to fully exploit the teacher model. To address these issues, we propose a method called Dense-Label Refinement and Cross Optimization (DRCO) based on dense pseudo-labels. Specifically, to tackle the issue of low-quality sampling, we introduce the Adaptive Sampling Approach (ASA), which achieves high-quality sampling at the image level and dynamic sampling ratios without introducing any additional hyperparameters. For the inconsistency between classification scores and localization quality, Comprehensive FRS (cFRS) is proposed to jointly optimize the classification and localization branches more efficiently, thereby obtaining a more comprehensive score. Finally, for the suboptimal training approach, we introduce Cross Prediction Optimization (CPO). CPO efficiently leverages the knowledge of the teacher model through a Cross-Head operation, thereby achieving more effective teacher-student interaction. DRCO achieves 27.31% mAP with only 1% COCO labeled data, which is approximately 1.24% mAP higher than the previous state-of-the-art. Based on MS COCO and PASCAL VOC benchmarks, further comprehensive experiments demonstrate that our method alleviates the aforementioned shortcomings and achieves competitive performance.
Keywords:
Optimization
Adaptation models
Training
Object detection
Predictive models
Location awareness
Data models
Detectors
Proposals
Image classification
Semi-supervised object detection
object detection
semi-supervised learning
Journal
IF:
3.6
Papers:
9.8W
Citations:
29.4W