Menlo Park with Nearmap AI
This video reel features an aerial imagery shot of the Menlo Park area in California, captured by Nearmap. Nearmap is a location intelligence company that designs and build the technology from cameras, to visual processing, and AI.
We cycle through layers relating to the urban environment, using the pixel probabilities of semantic segmentation output layers as transparency.
The handful of layers shown here depict the coexistence of urban infrastructure and the environment; roof, solar panels, gable roof, shingle roof, natural pervious surfaces, tree canopy, lawn grass, hard impervious surfaces, driveable surfaces, cars, stop markings, pedestrian crossings, and shadows.
80 individual semantic layers like these are produced as outputs by a single, highly customised deep learning model. Various tricks allow it to be trained and perform inference on very high resolution imagery, without seamlines on edges between tiles, and within practical GPU memory constraints.
The single deep learning model runs on millions of square kilometres of urban aerial imagery each year, ultimately producing ever updating, automated maps of the urban world in USA, Canada, Australia and New Zealand. Detail and scale form a single continuum, with national studies of urban tree canopy, changes to buildings and disaster response forming from pixels the size of playing cards.
