This paper presents a semi-supervised approach to training object detection systems using self-training. The authors aim to reduce the effort required to collect and label training data by combining a small set of fully labeled examples with a larger set of weakly labeled or unlabeled examples. The key contributions of the study include demonstrating that models trained with this method can achieve comparable performance to those trained with a larger set of fully labeled data, and that a training data selection metric independent of the detector outperforms metrics based on detection confidence. The paper begins by introducing the challenges of collecting and labeling training data for object detection systems, emphasizing the need for a streamlined approach. It then describes the semi-supervised training process, including the use of Expectation-Maximization (EM) and self-training methods. The authors focus on the incremental addition of training data, where the model is updated iteratively based on the confidence of weakly labeled examples. The experimental setup involves using a detector trained with a limited set of fully labeled examples and a large set of negative examples. The semi-supervised training procedure includes normalizing and generating synthetic training examples, running the detector over weakly labeled data, labeling and scoring weakly labeled examples, and iteratively adding the most promising examples to the training set. The selection metrics used are compared, with the detector-independent Mean Squared Error (MSE) metric outperforming the confidence-based metric. The experiments show that the addition of weakly labeled data significantly improves detector performance, even with a small initial labeled training set. The paper also discusses the importance of detector accuracy in localization and the impact of the initial training set size on performance. Finally, the paper concludes with observations on the feasibility of using weakly labeled data with existing detectors and the advantages of the MSE selection metric. It highlights the need for further research on practical applications, including the use of different detector versions, the relationship with co-training approaches, and the selection of initial training sets.This paper presents a semi-supervised approach to training object detection systems using self-training. The authors aim to reduce the effort required to collect and label training data by combining a small set of fully labeled examples with a larger set of weakly labeled or unlabeled examples. The key contributions of the study include demonstrating that models trained with this method can achieve comparable performance to those trained with a larger set of fully labeled data, and that a training data selection metric independent of the detector outperforms metrics based on detection confidence. The paper begins by introducing the challenges of collecting and labeling training data for object detection systems, emphasizing the need for a streamlined approach. It then describes the semi-supervised training process, including the use of Expectation-Maximization (EM) and self-training methods. The authors focus on the incremental addition of training data, where the model is updated iteratively based on the confidence of weakly labeled examples. The experimental setup involves using a detector trained with a limited set of fully labeled examples and a large set of negative examples. The semi-supervised training procedure includes normalizing and generating synthetic training examples, running the detector over weakly labeled data, labeling and scoring weakly labeled examples, and iteratively adding the most promising examples to the training set. The selection metrics used are compared, with the detector-independent Mean Squared Error (MSE) metric outperforming the confidence-based metric. The experiments show that the addition of weakly labeled data significantly improves detector performance, even with a small initial labeled training set. The paper also discusses the importance of detector accuracy in localization and the impact of the initial training set size on performance. Finally, the paper concludes with observations on the feasibility of using weakly labeled data with existing detectors and the advantages of the MSE selection metric. It highlights the need for further research on practical applications, including the use of different detector versions, the relationship with co-training approaches, and the selection of initial training sets.