众包全景拼接系统下基于图像质量的酬劳分配策略

Quality-Driven Reward Distribution in Crowdsourced Panoramic Map Construction

大量高质量的图像数据是全景地图构建的基本需求.然而,传统的街景图像采集方法大多需要雇佣专业的采集人员和配备专门的采集设备,这是耗时耗力且昂贵的.移动众包这一新型工作模式的发展与广泛运用启发我们借助群体智慧的力量共同采集街景图像.然而,任务中可能面临的资源损耗与参与者的利己主义心理使得他们大多并不乐意无偿地付出劳动.因此,我们考虑采用金钱激励措施来调动参与者的感知积极性并吸引他们加入感知任务.此外,由于参与者的个人能力、专业度等个人因素的差异性导致采集的街景图像质量参差不齐.考虑到大量低质量图像的掺杂会降低全景地图的构建性能,因此,我们将街景图像质量与参与者酬劳相结合,提出基于图像质量的酬劳分配机制,鼓励参与者采集高质量的、对全景构建高贡献度的图像.我们提出了基于权值和基于夏普利值的酬劳分配机制,分别命名为WPM(Weighted-based Pricing Mechanism)和SVPM(Shapley Value-based Pricing Mechanism).我们通过提取图像的SIFT(Scale-Invariant Feature Transform)特征值来表征图像内容,并通过量化图像对全景地图构建(即有效SIFT特征)的贡献度来确定图像质量.贡献度越大意味着图像质量越高.具体来说,在WPM中,单张图片质量被定义为该图中所有SIFT特征的价值之和.在SVPM中,图像质量被定义为该图片在不同街景拼接组合下的边际贡献的平均值.最后,平台根据图像质量给参与者分配酬劳.在实验部分,我们将提出的酬劳分配机制与传统分配方法进行性能对比.结果表明,我们的机制具有鲁棒性,并在公平性方面优于传统方法.

A large collection of high-quality images are the basic and indispensable demands for the construction of panoramic maps.However,the traditional image acquisition methods depend on hired professional staff and specialized equipment,which are expensive as well as time-consuming.The development and widespread of mobile crowdsourcing inspire us to collect diversified street view images by the wisdom of crowds.However,the potential costs and self-interests of participants make them unwilling to work without any profit.Thus,we consider motivating them to join crowdsourcing tasks through some reasonable financial incentives.Besides,due to the differences in personal capabilities and professionalism,the quality of collected images is different.High-quality images are beneficial to construct a complete and accurate panoramic map,while low-quality images may degrade the performance of panoramic maps.Thus,we combine image quality with the rewards of participants,design image quality-based reward distribution mechanisms,and encourage participants to collect high-quality images.We propose a weighted-based and a Shapley value-based pricing mechanism,named WPM and SVPM respectively.Besides,we extract SIFT to represent images.The image quality is quantified as the contributions of images to the construction of panoramic maps,i.e.,effective SIFT features.The higher contribution means a higher-quality image and brings a higher reward.Specifically,in WPM and SVPM,we quantify image quality based on different design principles,for example,in WPM,the quality of a single image is quantified as the valuation of its total effective features.Finally,we evaluate the performance of reward distribution mechanisms through extensive experiments.The results demonstrate that our mechanisms are superior to some traditional ways,like averaging and VCG,in terms of fairness and robustness.