A Breeding Animal Detection Method from Images With Grided Heatmaps

In practical scenarios for detecting objects in images where we cannot collect sufficient effective annotated data, we propose a new method that can utilize pre-trained classification network models to implement object detection tasks. Our method generates multiple scale gridded heatmaps from a pre-trained classification network, which can effectively integrate various pre-trained features of different categories without training. We also utilize similarity segmentation results to further refine the object detection performance. Our experiment results on COCO, VOC and our datasets show that our method can produce better detection results compared to some existing object detection networks when the number of training data is small for other networks.