Project Background

The client company had to spend a lot of time identifying the materials, equipment, and other products depicted in the architectural drawings (hereafter referred to as "picking out"). This process also requires expertise, and the accuracy of the process varies from one operator to another.

Therefore, a PoC project utilizing the rapidly evolving AI was considered to improve the efficiency and accuracy of the process of picking out products from architectural drawings, and to study the issues involved in putting the technology to practical use.

On the other hand, there are many solution vendors that have experience in AI technology for recognizing and inferring objects from image data, and the company needed assistance in selecting the most suitable vendor for this PoC project, given the lack of internal resources.

Process

Project Planning Phase
JP Tokyo organized the requirements for this PoC project through interviews and discussions with the client company and prepared the RFI (Request for Information), RFP (Request for Proposal), and evaluation criteria. JP Tokyo facilitated a Q&A session and a proposal presentation with each solution vendor and compiled the final selection.

• RFI target: Research and develop a short list of just over 20 vendors that specialize in image processing and recognition using AI. Reviewed with client companies and narrowed the list to just under 10 companies.

• RFP target: Reviewed responses from vendors and narrowed down to a few potential vendors for the RFP.

Project Implementation Phase
While the selected vendor was to facilitate the project with the client company, JP Tokyo took on the role of PMO in an advisory capacity. Based on the skills and experience of JP Tokyo's consultants, we assisted the vendor in ensuring that the project was proceeding in accordance with its objectives and that the project evaluation report was understandable to the readers, especially the project sponsor and other relevant department members.

Results

This PoC project had two major objectives, and the objectives were achieved almost as expected. The first objective was to confirm the accuracy of drawing recognition and inference. Inference on learned vector plans (drawing represented by line segments. Converted to jpeg when inputted) confirmed that increasing the number of drawing variations enabled highly accurate inference (high FIT rate: rate depicting how accurately the inference was made, high reproduction rate: rate depicting how well the inference was made without missing the details, low counterfactual error rate: rate depicting how many false conversions were made).

Another objective was to acquire basic knowledge on screen recognition and inference with AI. In the field of machine learning image recognition and inference, we were able to acquire basic technical knowledge of required tasks and work volumes, as well as knowledge on basic evaluation methods regarding image recognition and optical character recognition/reader analysis.

While the solution vendor prepared the project evaluation report, JP Tokyo prepared an evaluation report for executives and was able to report the results of the PoC project to a wide range of readers.