A CAM-Guided Parameter-Free Attention Network for Person Re-Identification

Most existing attention mechanisms have no supervised signal during the training phase, which limits model feature learning capability. To solve this problem, we propose a novel parameter-free mechanism based on class activation mapping. Attention usually consist of spatial and channel attention, indicates that “where” “what” is more meaningful, respectively. Our also contains both types attention. For Spatial Attention, use mapping as supervision to guide generation it directly in space. Thus our can pay informative pedestrian parts scene reduce background interference. Channel importance each obtained by similarity between aforementioned map channel. In manner, indirectly guided addition, parameter-free, reduces risk over-fitting. Finally, conduct extensive evaluations three popular benchmark datasets including Market1501, DukeMTMC-reID, MSMT17, demonstrating effectiveness approach discriminative person representations.