Segmentation-based bounding box generation for omnidirectional pedestrian detection.

We propose a segmentation-based bounding box generation method for omnidirectional pedestrian detection that enables detectors to tightly fit bounding boxes to pedestrians without omnidirectional images for training. Due to the wide angle of view, omnidirectional cameras are more cost-effective than standard cameras and hence, suitable for large-scale monitoring. However, state-of-the-art pedestrian detectors cannot directly be applied to omnidirectional pedestrian detection because pedestrians' appearance in omnidirectional images may be rotated to any angle, which substantially degrades the performance of standard pedestrian detectors. Existing methods mitigate this issue by transforming images during inference, though the methods have limited detection accuracy and slow detection speed. A recently proposed method obviates the transformation at the cost of laborious annotation works and trains detectors with omnidirectional pedestrian detection datasets. We propose instead leveraging an existing large-scale object detection dataset to obviate both the transformation and annotation works. We train a detector with rotated images and tightly fitted bounding box annotations generated from the segmentation annotations in the object detection dataset, which enable detectors to detect pedestrians in omnidirectional images with tightly fitted bounding boxes. We also develop pseudo-fisheye distortion augmentation, which deforms images of perspective view to imitate the fisheye distortion and further enhances the performance. Extensive analysis shows that our detector successfully fits bounding boxes to pedestrians and demonstrates substantial performance improvement over existing methods. [ABSTRACT FROM AUTHOR]