'Parallel-Circuitized' distillation for dense object detection

Yaoye Song, Peng Zhang, Wei Huang, Yufei Zha, Tao You, Yanning Zhang

Abstract: Highlights•A parallel knowledge distillation framework for dense object detection is proposed, which is inspired from analogizing the principle of parallel circuit to enhance effect of the dual-stream structured networks.•A Soft Distribute-Guided Quality Predictor (SDGQP) is proposed based on Softmax for dynamical selection of distribution statistics, which enables more reliable Localization Quality Estimation (LQE) for detection.•A localization quality distillation is also proposed based on more reliable localization quality score of SDGQP to bridge the gap between classification and bounding box regression branch.