Google AI - Object Segmentation

Around 2022-23 I had the opportunity to work with the Vision AI development team at google on some model training to support a few Behr digital visualization initiatives. The objective was to trained a model to identify paintable surfaces on uploaded images in less than ideal uploaded images. Extracted UGC from Behr’s Paint Your Place digital property from the past 12 years were selected and sorted masked with the objective to identify wall borders. These images included low res photos from older digital cameras with poor lighting conditions and soft focus. We made several batches of segmentation masks on a variety of images. It was interesting to see the results after each training session on what things regressed or became more accurate over time with trained data. The experience was an enlightening one where Google gave us access and taught us how to use their system but then gave us the freedom to explore and try different techniques to try and achieve our desired goals.

Behr Google AI

Visual Inspection AI Tool data set and prediction example
The dashboard to the AI model shows the labeled data set (above). Below is a prediction result example using one of the Behr blog images. You can see that the image is considerably dark, has a sheer drapery, a blurry glow from the table light as well as difficult areas around the plants. The gold color represents the predicted “wall” area mask not a colorization layer.

Behr Google AI