DDBN: Dual detection branch network for semantic diversity predictions

Highlights：

• A brand-new detection network, the Dual Detection Branch Network (DDBN), is the first exploration of multiple predictions (more than one prediction) in the field of object detection. Experiments have also proved that semantic diversity prediction (multiple predictions) can indeed bring accuracy improvements.

• A specialized training method, i.e., Diversity Enhancement Strategy (DES). Because our DDBN is the first multi-prediction framework in the object detection field, our model trained with the traditional direct regression method does not improve the detection performance significantly, so we customize an effective training method (DES) for the first multi-prediction model.

• Our DDBN has brought significant accuracy improvements on multiple benchmark datasets, which shows the generality and superiority of our model.

摘要

•An effective feature fusion method (Adjacent Feature Compensation (AFC)) that generates different semantics for object detection. Although the structure is somewhat similar to the existing methods of the top-down and the bottom-up, it is indeed the first method to generate different semantic features in the field of object detection.•A brand-new detection network, the Dual Detection Branch Network (DDBN), is the first exploration of multiple predictions (more than one prediction) in the field of object detection. Experiments have also proved that semantic diversity prediction (multiple predictions) can indeed bring accuracy improvements.•A specialized training method, i.e., Diversity Enhancement Strategy (DES). Because our DDBN is the first multi-prediction framework in the object detection field, our model trained with the traditional direct regression method does not improve the detection performance significantly, so we customize an effective training method (DES) for the first multi-prediction model.•Our DDBN has brought significant accuracy improvements on multiple benchmark datasets, which shows the generality and superiority of our model.