SINGLE-SHOT DETECTOR USING ADAPTIVE ANCHORS

Slim-object detection is one of the challenging problems in image processing because the shape (or bounding box) of a slim object changes a lot according to the viewpoint. However, so far there has been a lot of investigations on small-object detection problems. In addition, most of investigations were focused on effective feature extractions. However, only with the effective feature extraction slim-object detection problems are not manipulated properly because of the large-scale varying proportions of bounding boxes. In general, most of single-shot detectors use anchors to detect several objects at a grid cell. However, those grid anchors are not distributed properly. In order to make the best use of anchors, a new anchor (called adaptive anchor) is proposed in this paper. The major difference between grid anchors and adaptive anchors is that the strides of adaptive anchors do not depend upon the shapes of feature maps. In order to estimate the efficiency of adaptive anchors we trained tie detection models (using grid anchors and adaptive anchors). Training results shows that the adaptive anchors are more suitable than grid anchors for slim-object detections.