Focus, Align, and Sustain: Counteracting Gradient Dilution in Incremental Object Detection
Researchers have introduced a new framework called FAS to address gradient dilution in incremental object detection using Detection Transformers. This phenomenon, where optimization signals weaken over sequential learning, is caused by signal dispersion, assignment drift, and support attrition. FAS aims to counteract these issues by focusing gradient flow, aligning query-target assignments, and sustaining the feature space support of older classes. AI
IMPACT This research offers a novel approach to improve the performance of object detection models in sequential learning scenarios.