Research On 3D Body Reconstruction For Ordinary Users

Due to the needs of games,animation and other applications,3D human reconstruction has always been an important research topic in computer graphics.In recent decades,re-searchers have proposed a variety of methods around this topic,and have many achievements.Since human body model has high application values in games and animation,these methods can be tolerated to pay a high cost in the aspects of scanning equipment,requirements and reconstruction efficiency.However,for individual users of virtual fitting,these methods have many shortcomings in terms of convenience and efficiency as follows:(1)scanning devices such as scanners and Kinect are not commonly used by ordinary users;(2)current ways require users to wear tight or little clothes,which brings inconvenience and embarrassment to users;and(3)iteratively optimizing the human body model to fit the input data is very time-consuming.This thesis overcomes several difficulties in terms of improving convenience and efficien-cy.After three stages of research,we ended up using only one Android phone or tablet to take two photos of the user in front and side views,and then obtain the user's 3D body in seconds.In order to facilitate the user to further edit the 3D body shape,we propose a parametric semantic model.Specifically,the contributions of this thesis are summarized as follows:·We propose a fast clothing size prediction method using dressed-human images.Dressed-human images refer to the images where human wear common clothes.With dressed-human images as input,on one hand,users can stay away from the scanners or depth sensing devices that are not widely available,and on the other hand,they avoid the inconvenience and embarrassment caused by wearing minimally.Existing methods have difficulties in effectively and efficiently estimate body shapes under clothes.In order to effectively remove clothes occlusions,we construct a database containing 3 sets of common clothes,a total of 6042 × 3 3D undressed and dressed body pairs.The database is of great significance for the research of human body estimation and cloth simulation.In order to efficiently estimate body sizes,we design a set of body land-marks as a bridge connecting the dressed-human image and body sizes.Based on the constructed database,we propose a data-driven method for efficiently recommending clothing size.Firstly,we regress body landmarks from input image,and then learn body sizes using landmarks.Finally,according to body sizes,we acquire clothes size through automatically querying the clothes size table.The experimental results show that our method can effectively and efficiently remove the influence of clothes,and then estimate body sizes under the cover of clothes.·We propose a 3D human reconstruction method using dressed-human images.When shopping online,to help users make decisions and increase the rate of successful transac-tions,we not only need accurate clothing size recommendations,but also need reliable visual effects of dressing clothes.This requires us to rebuild a 3D human body model that is consistent with the user's body.In view of the difficulties introduced by dressed-human image,we adopt the solution described in the above paragraph which makes use of database and body landmarks.In contrast,in order to reconstruct 3D human body,we first regress 3D landmarks,and then use landmarks as constraints to optimize the parameters of human body model.To more effectively describe the relationship between body landmarks and dressed-human silhouette,we propose a landmark-based feature descriptor,which improves the result of landmarks regression.We also verify the key configurations of implementation through rich comparative experiments,which further improves the results.Under the premise of ensuring reconstruction accuracy,our method has greatly improved the efficiency of reconstruction.The process of our 3D landmarks regression and 3D body reconstruction takes less than 4 seconds,while the current fastest method for 3D body reconstruction under clothes needs 1 minute.·We propose a convenient and efficient 3D human reconstruction method for ordinary users.The previous two phases of work are all running on the PC,and additional segmentation tools(such as Photoshop)are needed to generate silhouette as input,which brings inconvenience to the user.Therefore,we develop a user-friendly 3D body reconstruction method based on mobile devices.Due to the limited resources and computational power of mobile devices compared to PCs,we need more efficient body reconstruction methods.Based on the constructed database,we learn body parameters directly from dressed-human silhouettes,instead of iteratively optimizing body parameters with landmarks as constraints.The parameters regression process takes about 1.26 seconds on an ordinary Android phone.In order to directly use the photos taken by the mobile phone,we develop a user-friendly interactive tool for image segmentation to extract the silhouette on the mobile terminal.In addition,we design a new strategy to avoid camera calibration,which further improves convenience.·We propose a semantic parametric human body model,based on which,we reconstruct and edit 3D body shapes according to several body measurements.The parameters of existing body models cannot perform semantic controls of body shapes,where semantic controls refer to the intuitive deformations such as "thinning waist" and "lengthening legs".We divide the template body into several parts according to body sizes.Based on the human body database,we analyze the shape variations of each body part along length and girth direction to learn semantic bases.In order to express the body shape space more completely,we also learn non-semantic bases in addition to the semantic bases.We only use the semantic bases to represent the bodies in the training database,and then calculate the difference between the represented ones and the corresponding ground truth.We acquire the non-semantic bases by analyzing the differences.The semantic bases enhance the model's ability to express local deformations,while the non-semantic bases guarantee the coherence between neighboring body parts.We also learn the mapping between anthropometric measurements and body parameters to reconstruct or edit body shapes according to measurements.To bring users more convenience,we allow any number of measurement as input,and use a correlation-based method to predict unknown measurements with known measurements.The experimental results show that our method reduces the estimation error of unknown measurements,and demonstrate the effectiveness of the model in human reconstruction and editing aspects.