On the Arbitrary-Oriented Object Detection: Classification Based Approaches Revisited

Arbitrary-oriented object detection has been a building block for rotation sensitive tasks. We first show that the boundary problem suffered in existing dominant regression-based detectors, is caused by angular periodicity or corner ordering, according to parameterization protocol. also root cause ideal predictions can be out of defined range. Accordingly, we transform prediction task from regression classification one. For resulting circularly distributed angle problem, devise Circular Smooth Label technique handle and increase error tolerance adjacent angles. To reduce excessive model parameters Label, further design Densely Coded Labels, which greatly reduces length encoding. Finally, develop an heading module, useful when exact orientation information needed e.g. ship plane detection. release our OHD-SJTU dataset OHDet detector Extensive experimental results on three large-scale public datasets aerial images i.e. DOTA, HRSC2016, OHD-SJTU, face FDDB, as well scene text ICDAR2015 MLT, effectiveness approach.