I rewrote "Deep Floor Plan Recognition Using a Multi-Task Network with Room-Boundary-Guided Attention" (hereinafter referred to as "Deep Floor Plan") announced at ICCV2019 to the python3 version and tried it. [paper, github python3 version)](https://github.com/c-makitahiroki/DeepFloorplan)]
I was worried that the task might have specific business needs. In this article, I will briefly introduce DeepFloorPlan and then write the result of running the demo program. The results of running the learning program will be added in the future.
In a nutshell, it's a DNN that "** recognizes the room layout **". It is characteristic that the decoder is separated for each room boundary prediction task and room type prediction task. In addition, room boundary information is used as an attention for room type prediction.
The network itself is relatively simple, but better performance than PSPNet and DeepLab V3 + Is out.
OS: windows10 64bit GPU: GeForce GTX1060 6GB
I tried running the demo program using the trained model. Here, I would like to give an example where reasoning is relatively good and an example where it is not.
Some of the halls are recognized as closets and toilets, but the background and room types, as well as the doors and windows, are relatively correctly recognized.
The boundary between the balcony and the dining room is not recognized correctly. It seems that the door type is a little special. Also, this data does not recognize some of the background and room boundaries correctly. It seems that the cause is that the annotation of the drawing is included in the background part of the input image.
I looked at some other data, but the following points seem to be difficult.
I would like to make improvements to deal with this area in the future.
I would like to improve it and add the results later. If you have any suggestions for improving accuracy, please comment!
Recommended Posts