01
The paperHuang & Zheng · 2018
- Pix2PixHD used to generate zoning maps of floor plans
- Detail improves with number of epochs - ResNet used to prevent overfitting
- CGAN - Conditional Generative Adversarial Networks - punish trining going in wrong direction
- Pix2PixHD - 5 groups of convolutional layers, 9 groups of residual network layers, 5 groups of deconvolutional layers
- Using PyTorch - images from lianjia.com
- Labels used: Walkway, Bedroom, Living Room, Kitchen, Toilet, Dining Room, Balcony, Window, Door
- 115 image pairs - 100 training, 15 testing ~ 2hrs to train one network
- Works well in recognising bedroom, kitchen, toilet & balcony - normally well-defined by walls and specific furniture
- Less successful for walkway/living room - but some architects got this wrong when labelling test data (done by Architectural students)
- Struggled with non-orthogonal shapes
- In one image, GAN found mis-labelled data. Later found 4 errors in human-labelled data, which didn't lead to training in the wrong direction - pretty robust network.
- Generating plans from zones - sometimes put furniture facing the wrong way. Some architects agreed with the network's furniture position above the position in the test images.
- 2-pixel convolutions. Residual network avoids overfitting.
- Trained to 80 epochs - loss value stabilised after that
- The paper looks at individual neurons' feature detection
- Bedrooms/balconies often activated together
We help professional services firms work out where AI genuinely helps, what has to be structured first, and what to leave alone.
Get in touch