保持高分辨率信息的无锚点框检测算法

目标检测逐渐成为视觉研究社区的关键领域,而其挑战之一是检测器难以准确地定位不同尺度的物体.面向图像中的目标检测应用,提出了高分辨率-无锚点框(HOAR)检测策略来应对物体尺度多变的挑战.HOAR将待测图像输入多条通路(对应不同尺度)并行的高分辨率网络,并提取每条通路上的输出特征图作为图像在每种尺度下的深度特征表示;然后...

High Resolution Information Reserved Anchor-Free Detection Algorithm

Recently,object detection has gradually become a critical domain in the visual research community,while it’s also challenging to locate objects of different scales accurately for detectors.For the application of detecting objects in images,a high resolution anchor free(HOAR)detection strategy is proposed to meet the challenge of variable object scales.HOAR first inputs the image into a parallel high-resolution network with multiple paths(which is corresponding to different scales).Then the output feature map is extracted from each path as the deep feature representation of the image at each scale.Next,the dense feature pyramid(DenseFPN)is adopted to fuse the information of these feature maps to obtain the re-combined feature maps of multiple scales.Finally,the anchor-free detection sub-network is used to determine the object’s category and rectangle box position of each point on these feature maps.To verify the effectiveness of HOAR,comparative experiments are carried out on COCO dataset.First,the results on ablation study show the necessity of each module in the HOAR strategy.Second,the evaluation metric of detection,namely mAP,of HOAR reaches 40.5 on the validation set,which is significantly higher than mAP of all baseline models and some SOTA methods.In addition,the size of model parameters of HOAR strategy is significantly less than that of baseline models.