What is Adaptive Training Sample Selection (ATSS)?
Adaptive Training Sample Selection (ATSS) is a method that selects positive and negative samples by analyzing the statistical characteristics of an object. It combines the anchor-based and anchor-free detectors in computer vision to improve object detection models.
How does ATSS work?
ATSS selects positive samples by finding the candidate samples based on the center of the ground-truth box on each pyramid level. The number of candidate samples is determined by a constant variable K.
The intersection-over-union (IoU) metric is used to determine if the candidate samples qualify as positive samples. The mean and standard deviation of the IoU values between the candidate samples and the ground-truth box are calculated, which documents the level of similarity between them. This information determines the threshold of the IoU values to be used to qualify a candidate sample.
If a candidate sample’s IoU is greater than or equal to this threshold, it is considered a final positive sample. ATSS also limits the positive samples' center to the ground-truth box, ensuring that they are only selected if they are inside it. If an anchor box is assigned to multiple ground-truth boxes, the one with the highest IoU will be selected as positive. The remaining anchor boxes are negative samples.
Why is ATSS Important for Object Detection Models?
ATSS has certain advantages that classical object detection models do not offer. It reduces the impact of some ill-aligned anchors. It improves the limit of object proposal methods and is more suitable for high-density object detection scenarios. ATSS bridges the gap between anchor-based and anchor-free detectors, making object detection models more efficient and accurate. It can also be helpful in the detection of objects with irregular shapes, such as pedestrians or animals.
Adaptive Training Sample Selection (ATSS) is a promising method for improving object detection models in the field of computer vision. It uses statistical characteristics of objects to select positive and negative samples to bridge the gap between anchor-based and anchor-free detectors. ATSS will lead to more efficient and accurate object detection, especially in high-density or irregular shaped object detection scenarios.
