Inference Overlay: Difference between revisions

From Tygron Preview Support Wiki
Jump to navigation Jump to search
No edit summary
No edit summary
 
(34 intermediate revisions by 2 users not shown)
Line 1: Line 1:
The AI Inference Overlay is a [[Grid Overlay]] which can spatially identify features using one or more Prequel Grids. Features are identified using a [[Neural Network]]. This neural network takes a subsection (window) of the input grid and either classifies or detects one or more objects in that window. The window of detection marches over the input grid with a configurable stride.  
[[File:Treesmask.jpg|thumb|right|Trees detected with Inference Overlay.]]
The AI Inference Overlay is a [[Grid Overlay]] which can spatially identify features using one or more Prequel Grids. Features are identified using a Convolution [[Neural Network]] that used a limited amount of proposal regions (RCNN). This Neural Network takes a subsection (window) of the input grid and either;
* [[Neural Network (Inference Overlay)#Supported Convolution Types|classifies]] the image as a whole
* [[Neural Network (Inference Overlay)#Supported Convolution Types|detect one or more features]] in that image.  


It is not possible to run the convolution network on the grid as a whole. Therefore, a marching window strategy is applied instead, where the window is moved over the input grid with a [[Stride fraction (Inference Overlay)|configurable stride]]. This window takes the input values of one or more grids and puts these into tensors representing either floating point values or color channels (red, green and blue). Depending on the dataset used to train the neural network, the tensor values are normalized based on a particular value range. For example, a relative height input tensor for a foliage neural network trained on height data can be normalized using a range of 0 to 40 meters, when 40 meters is considered the largest tree height in the Netherlands.


==Foliage Example==
[[File:inference_foliage_animated.gif|frame|right|Animation of inference with a moving window and Bounding Box detection and a [[Stride fraction (Inference Overlay)|stride fraction]] of 0.5.]]
Using a [[Satellite Overlay]] of 0.1m detail, foliage features can be identified using an [[Inference Overlay]] and enhanced with a [[Digital Terrain Model Overlay]] (DTM), a [[WCS Overlay|WCS Overlay]] representing the [[DSM]], [[Combo Overlay]] to combine these and an optionally an iterative [[Avg & interpolation (overlay)|Max Overlay]] to enhance the foliage height. For more information, see this [[How to detect foliage using an Inference Overlay|how-to]].
<gallery widths="200">
foliage_inference_labels.jpg|[[Labels result type (Inference Overlay)|Labeled features]] on 0.1m satellite image
foliage_inference_scores.jpg|[[Scores result type (Inference Overlay)|Label Scores]]
foliage_inference_masks.jpg|[[Masks result type (Inference Overlay)|Pixel Masks]]
foliage_inference_boxes.jpg|[[Boxes result type (Inference Overlay)|Bounding Boxes]]
foliage_inference_foliage_height.jpg|[[DSM]] subtracted by [[DTM]] on identified features
foliage_inference_iterative_max_5_iterations_d0_25m.jpg|Max neighboring height within 0.25m, iterated 5 times
</gallery>


{{article end
{{article end
Line 9: Line 24:
* [[ONNX]]
* [[ONNX]]
* [[PyTorch]]
* [[PyTorch]]
|howtos=
* [[How to detect foliage using an Inference Overlay]]
* [[How to create foliage height based on an Inference Overlay]]
* [[How to import trees based on an Inference Overlay]]
* [[How to detect solar panels using an Inference Overlay]]
* [[How to update Buildings's solar panel attribute based on an Inference Overlay]]
}}
}}
{{InferenceOverlay nav}}
{{InferenceOverlay nav}}
{{Overlay nav}}
{{Overlay nav}}

Latest revision as of 11:19, 19 December 2024

Trees detected with Inference Overlay.

The AI Inference Overlay is a Grid Overlay which can spatially identify features using one or more Prequel Grids. Features are identified using a Convolution Neural Network that used a limited amount of proposal regions (RCNN). This Neural Network takes a subsection (window) of the input grid and either;

It is not possible to run the convolution network on the grid as a whole. Therefore, a marching window strategy is applied instead, where the window is moved over the input grid with a configurable stride. This window takes the input values of one or more grids and puts these into tensors representing either floating point values or color channels (red, green and blue). Depending on the dataset used to train the neural network, the tensor values are normalized based on a particular value range. For example, a relative height input tensor for a foliage neural network trained on height data can be normalized using a range of 0 to 40 meters, when 40 meters is considered the largest tree height in the Netherlands.

Foliage Example

Animation of inference with a moving window and Bounding Box detection and a stride fraction of 0.5.

Using a Satellite Overlay of 0.1m detail, foliage features can be identified using an Inference Overlay and enhanced with a Digital Terrain Model Overlay (DTM), a WCS Overlay representing the DSM, Combo Overlay to combine these and an optionally an iterative Max Overlay to enhance the foliage height. For more information, see this how-to.