AI-woordenlijst
Het complete woordenboek van kunstmatige intelligentie
Classification Subnet
RetinaNet subnet responsible for predicting the probability of object presence for each anchor, using 4 convolutional layers with 256 filters each.
Regression Subnet
Parallel subnet that predicts precise bounding box coordinates for each anchor, with an architecture identical to the classification subnet.
ResNet Backbone
Residual neural network used as the main feature extractor in RetinaNet, typically ResNet-50 or ResNet-101 to extract rich representations.
Alpha Parameter (γ)
Focal loss hyperparameter (typically γ=2) that controls the decay rate of the loss for well-classified samples, adjusting the focus on hard examples.
Balanced Focal Loss
Variant of focal loss that includes an α parameter to balance the weight between positive and negative classes, thus optimizing the detection of rare objects.
Pyramid Levels
Feature levels in FPN (P2 to P7) generated from different backbone layers, each capturing information at a specific resolution scale.
Box Encoding
Parameterization of bounding box coordinates using offsets relative to anchors rather than absolute coordinates, stabilizing network training.
Smooth L1 Loss
Loss function used for bounding box regression in RetinaNet, combining the advantages of L1 and L2 loss for better robustness to outliers.