#76
summarized by : kaikai zhao
HEAD: HEtero-Assists Distillation for Heterogeneous Object Detectors

どんな論文か?

investigate how to effectively conduct knowledge distillation among heterogeneous teacher-student pairs for detection tasks (difficulty is the semantic gap between different backbone features)

新規性

1) add a detection head with architecture homogeneous to the teacher head attached to the student; 2) distillation on heterogeneous detection head (classification branch);

結果

achieved significant improvement compared to current detection KD methods

その他(なぜ通ったか?等)

cnn, feature distillation