Visual search services and applications project no : P3016-161

• Main Author: Li, Sisi
• Other Authors: Yap Kim Hui
• Format: Final Year Project
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/74606
• Institution: Nanyang Technological University

Image Recognition (IR) has been a hot research and development topic over the past few decades. More specifically and recently, Mobile Visual Search (MVS) is gaining great popularity with the wider usage of mobile devices such as PDAs, smartphones, wearable devices and tablets etc. Applied to search engine applications, comparing to traditional keyword / text-based searching method, IR/MVS-based searching method provides users with greater flexibility and feasibility in many scenarios (e.g. in case where the user only has a picture of the interested object and unable to provide keyword). Moreover, it can drastically improve the searching accuracy by matching exactly the same item/info that users are looking for, saving the time cost for the users. Extensive academic/research work has been done on IR/MVS topic and some very genius and practical algorithms were developed. Even available in the commercial market, such algorithms have been providing services and applications to users, performing ever improving functions and user experiences.It is, therefore, of this project’s interest to look into one of the most prominent existing algorithms, to study and to test its performance and functioning in some very tricky and interesting scenarios and set-ups. For the past projects with similar objectives, such work has been performed on items like painting, snacks etc. Whereas the algorithm has been tested with some quite positive results, extending the test to wider range of real-life applications scenarios, there are still some very unique set-ups to be explored. Particularly for this project, it focuses on T-shirts, representing objects with relatively loose shape (as it can be easily deformed to conform to the figure of the fitting person), or put it in another way – 3D objects with deformable surfaces.To achieve the objective of testing and evaluating MVS algorithms performances, the following steps were designed and executed with the following step:1. To build a small database containing images of fitted T-shirts, which bear unique pattern at its front side. To emulate real-life scenarios, all images were taken with author’s smartphone – iPhone 7 plus, one of the most popular models available in market now;2. To group images in the data base with principle of control-variable set-ups, i.e. in between different groups, there is one-variable change; 3.To run the algorithms with the built T-shirt image database with “reference set” and “test set”;4.To compare and evaluate on the accuracy of different methods 5.To introduce algorithm with Geometric Verification (GV) method and run the “test set” again; 6. To compare the GV results with those in step 3; Though having conducted the test with “control variable” set-ups, it is not possible to have a database can be exhaust all scenarios. Therefore, moving forward, the topics of improving image matching accuracy and optimizing database will be discussed as well.