Artificial Dummies for Urban Dataset Augmentation

Existing datasets for training pedestrian detectors in images suffer from limited appearance and pose variation. The most challenging scenarios are rarely included because they too difficult to capture due safety reasons, or very unlikely happen. strict requirements assisted autonomous driving applications call an extra high detection accuracy also these rare situations. Having the ability generate people arbitrary poses, with appearances embedded different background scenes varying illumination weather conditions, is a crucial component development testing of such applications. contributions this paper three-fold. First, we describe augmentation method controlled synthesis urban containing people, thus producing never-seen This achieved data generator (called DummyNet) disentangled control pose, appearance, target scene. Second, proposed relies on novel network architecture associated loss that takes into account segmentation foreground person its composition Finally, demonstrate generated by our DummyNet improve performance several existing across various as well situations, night-time where only amount available. In setup day-time available, detector 17% log-average miss rate over trained only.