Table 1.

Unified configuration for model training.

Config item	Parameter setting	Description
Maximum Iterations	40, 000	Validation is performed every 4, 000 iterations.
Optimizer	SGD	Learning rate: 0.0025, Momentum: 0.9, Weight decay: 0.0005
	AdamW	Learning rate: 0.0001, Beta1: 0.9, Beta2: 0.999, Weight decay: 0.01
Learning Rate Schedule	Polynomial decay (PolyLR)	SDG: 1e-4 AdamW: 1e-6
Pre-trained Weights	ResNet-101	The backbone network is initialized with weights pre-trained on ImageNet.
Training Resume	Checkpoint resume mode enabled	Automatically loads the best saved weights to continue training after an interruption.