研究目的
Investigating the effectiveness of a semi-supervised learning approach for object detection using weak annotation and a relative quality metric to generate improved detectors.
研究成果
The proposed semi-supervised approach for object detection using weak annotation and a relative quality metric successfully generates improved detectors, as evidenced by significant gains in recall with a small loss in precision. The method is particularly effective in challenging conditions such as rainy weather and low resolution images.
研究不足
The method requires a preexisting weak detector and involves a trade-off between precision and recall. The quality control step, while minimizing manual annotation, still requires some manual inspection.
1：Experimental Design and Method Selection:
The proposed method involves using detections from a preexisting weak detector (DPM) to train a new strong detector (RCNN). A quality control step is performed to evaluate the performance of both detectors.
2：Sample Selection and Data Sources:
A large dataset of streaming from public traffic cameras was collected, split into training and test sets, and further divided into uncontrolled weather and rainy subsets.
3：List of Experimental Equipment and Materials:
Intel CPU E5-2670 2.30GHz, 256GB RAM, GeForce Titan X.
4：30GHz, 256GB RAM, GeForce Titan X.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The weak detector (DPM) is applied to the training set, and the resulting detections are used to train the strong detector (RCNN). Quality control is performed on both detectors using a sample of the test set.
5：Data Analysis Methods:
The performance of the detectors is evaluated using precision and recall metrics, with relative gains calculated to assess improvement.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
