保持高分辨率信息的无锚点框检测算法

High Resolution Information Reserved Anchor-Free Detection Algorithm

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;然后利用密集特征金字塔(DenseFPN)对这些特征图进行信息融合,得到重新组合的多尺度特征图;最后采用无锚点框表示的检测子网络判断这些特征图上每个点对应的物体类别和矩形框位置.为验证所提HOAR策略的有效性,在COCO数据集上进行了对比实验.消融分析的结果表明了HOAR策略各个模块的必要性;其在验证集上的检测指标mAP达到了40.5,显著超过了基线模型和部分代表性算法的mAP.此外,HOAR策略所需的参数数量也显著小于对应的基线模型.

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.