This is an implementation of the Single Shot Object Detector architecture (Liu et al.). In particular, this project provides implementations of the Pooling Pyramid Network variation of SSD (Jin et al.), which achieves a meanAP comparable to the original version with the substantial reduction in parameter count.
This implementation proposes an architectural adaptation to allow training taking into account object subparts annotations.
The networks directory contains a few architecture variations. In particular, there are implementions alternating the base architecture, as well as containing or not our architectural tweak.
This project allows different SSD variations, simply by extending networks.ssd.SSD (without the subparts tweak) or networks.subparts_ssd.SubParts_SSD (to use the tweak), and simply following the examples of the other files.
The QR_Code_Experiments directory contains a series of experiments aiming at using the PPN version for detection of QR Codes in natural scenes. Here are a few exemple results of the best scoring model.
