Problem Description:
In target detection, the huge number of anchors causes a serious imbalance problem. What does imbalance mean here.
Questions and Answers:
In object detection tasks, Anchor refers to a set of predefined bounding boxes or candidate boxes that are combined in multiple scales and aspect ratios. The function of Anchor is to generate multiple suggestion boxes on the input image for use in the Region Proposal Network (RPN) in target detection or other similar tasks.
The imbalance problem means that in the target detection task, there is a large difference in the number of samples between positive examples (Anchor that contains the target) and negative examples (Anchor that does not include the target), that is, the proportion between positive examples and negative examples is unbalanced . Since the target usually only occupies a small area of the image, most anchors are negative examples, while the number of positive examples is relatively small.
This imbalance problem may cause the model to be biased towards learning negative examples during training, while insufficiently learning positive examples. Because the model only needs to predict negative examples in most cases, that is, most Anchors do not contain targets, if no measures are taken to solve the imbalance, the model may tend to produce a large number of negative example predictions, and the detection performance of positive examples will be poor. Poor.
To resolve the imbalance problem, methods commonly used include:
-
Sampling strategy: Properly sample positive and negative examples in the training data to balance their numbers.
-
Weighted loss: Weight the loss function of positive examples and negative examples so that the loss weight of positive examples is larger, thereby strengthening the learning of positive examples.
-
Online difficult sample mining: dynamically select those negative examples that are difficult to distinguish as training samples during training to improve the model's learning effect on negative examples.