Article Directory
- conceptual difference
-
- Difference 1: Positioning method
- Difference 2: Number of Anchor Points
- Difference 3: Feature Extraction
- Difference 4: Prediction Mode
- Difference 5: Applicable scenarios
- Difference 6: Inference Speed
- Difference 7: Model Complexity
- Difference 8: Difficulty level
- Difference 9: Processing Power
- Difference 10: Detection Accuracy
conceptual difference
Difference 1: Positioning method
The Anchor-based method uses a preset anchor point to determine the position and size of the candidate frame; the Anchor-free method does not need to set the anchor point in advance, and directly infers the object position and size from the feature map.
Difference 2: Number of Anchor Points
The Anchor-based method needs to manually set a series of anchor points to generate candidate boxes; the Anchor-free method does not need to set anchor points.
Difference 3: Feature Extraction
Anchor-based methods usually use a single convolutional layer to extract features, while anchor-free methods usually use multiple convolutional layers to capture features at different scales.
Difference 4: Prediction Mode
In the Anchor-based method, each anchor point is classified and regressed for a specific scale and aspect ratio; in the Anchor-free method, classification and regression can be performed for each pixel.
Difference 5: Applicable scenarios
The Anchor-based method is generally suitable for situations where the size of the object changes less and the number is relatively rare; the Anchor-free method is suitable for situations where the size of the object changes greatly and the number is denser.
Difference 6: Inference Speed
Since there is no need to generate a large number of candidate boxes, the anchor-free method is usually faster than the anchor-based method in terms of inference speed.
Difference 7: Model Complexity
The Anchor-based method requires additional operations (such as generating candidate boxes and calculating the IOU value of the real box, etc.), resulting in a relatively higher model complexity.
Difference 8: Difficulty level
The Anchor-based method needs to manually set the number and size of anchor points, which requires certain experience and skills; in contrast, the Anchor-free method is easier to implement and optimize.
Difference 9: Processing Power
In dense target areas, the detection results of the Anchor-free method may cause overlapping or repeated detection problems, while the Anchor-based method can deal with this situation by adjusting the position and size of the anchor points.
Difference 10: Detection Accuracy
In some cases, the Anchor-based method can achieve higher detection accuracy; and the Anchor-free method has better object scale robustness than the Anchor-based method.