Object Detection

Object detection is a task in computer vision, which requires the algorithm to predict a bounding box with a category label (class) for each region of interest (ROI) in an image.

What are anchor bozes?

Anchor Boxes are used to represent ideal shape and size of object the model predicts. Anchor boxes are pre determined boxes at different positions of an image with varying sizes and aspect ratio. These are determined by analysing the training data, where and how the most common annotations are present.

Advantages

• Enables predictor to detect the objects shape ,size and location for better detection

Disadvantages

• The model might not predict objects much smaller than the anchor box.
• The model might not predict objects much larger than the anchor box.

Anchored Object Detection

Anchored object detection algorithms uses 1000s of Anchor Boxes to determine regions of interest for classifying them as a potential candidate for an object. When more than one anchor box corresponds to an object, their IOU is taken and the label corresponding to highest score is selected.

Conventionally, anchor boxes are considered key for detectors.

Advantages

• Fastest method as it requires only a fixed number of boxes to be analysed.

Disadvantages

• The prediction is dependent on the size, number & aspect ratios of anchor boxes.

• Large number of anchor boxes for a small number of predictions

• Complex computation to calculate the IOU

Anchorless Object Detection

In an Anchorless model, every pixel in the feature map is predicted with an object box, similar to segmentation.

Advantages

• Detection framework becomes similar to segmentation & key point detection enabling us to reuse ideas

• Detection becomes anchor free which lowers the computation & design parameters

• There is no dependence on the size of the anchor box for detection

Disadvantages

• Models will take longer to converge