Limited supervision makes the task of pedestrian detection very challenging. Increasing the diversity of pedestrian samples is critical for semisupervised pedestrian detection (SPD). We propose to improve SPD from a new perspective by separating instance-specific content from scene style, thus performing content-based modification on the pedestrian samples collected from unannotated images of both target and external scenes. Toward this end, we propose a scene-adaptive and content-based instance modification model (SACIM), in which a pedestrian sample is represented by a scene-style code and a content code. One benefit of SACIM is that pedestrian samples from an external scene can be utilized by changing the corresponding scene-style codes. For the samples extracted from unannotated images, however, there is typically an issue of poor alignment. Another benefit is to address this issue by modifying their contents. The synthesized data will be used to improve pedestrian-background classification, which leads to more accurate pseudoannotations for training scene-specific pedestrian detectors. Extensive experiments with comparisons to the state-of-the-art SPD methods demonstrate the advantage of SACIM.
展开▼
