In object detection tasks, annotations are crucial for training machine learning models. However, obtaining high-quality annotations can be time-consuming and expensive. This paper explores the impact of noisy annotations on object detection models and proposes a novel approach to improve their performance in such scenarios.
Noisy Annotations and Their Impact
Annotations are the ground truth data used to train machine learning models. However, in real-world applications, annotated data can be noisy due to various reasons such as human error or inconsistent labeling. Noisy annotations can significantly affect the performance of object detection models, leading to decreased accuracy and increased false positives.
Proposed Approach: ATSS
To address the limitations of existing approaches, the authors propose a novel training data sampling method called Anchor-Free Training Strategy for Semi-Supervised (ATSS). This approach combines the strengths of anchor-based and anchor-free methods by proposing a novel training strategy that utilizes both types of anchors.
How ATSS Works
The proposed ATSS method works by selecting high-quality anchors from the noisy annotations and using them to train an object detection model. The anchors are selected based on their relevance to the object of interest, and the model is trained on a combination of clean and noisy data. This approach enables the model to learn more robust features that are less sensitive to noisy annotations.
Experimental Results
The authors evaluate the proposed ATSS method on several benchmark datasets and compare its performance with existing approaches. The results show that ATSS significantly outperforms existing methods in terms of accuracy and robustness against noisy annotations. Specifically, ATSS improves the AP50 score by 3.6% compared to the second-best approach.
Conclusion
In conclusion, this paper highlights the impact of noisy annotations on object detection models and proposes a novel approach to improve their performance in such scenarios. The proposed ATSS method combines the strengths of anchor-based and anchor-free methods and demonstrates superior performance compared to existing approaches. This work has important implications for real-world applications where annotated data can be noisy, and accurate object detection is critical.
