| --- |
| license: mit |
| tags: |
| - neural_network |
| - CNN |
| - training_data |
| --- |
| |
| # Collision |
|
|
| ## Intro |
| Collision is a simple proof of concept at https://github.com/osick/collision of a neural network detecting collsions of bodies (see the images above). In this setup it can reach an accuracy of about 98% which sounds good, but in fact is not good enough remembering self driving cars ... |
| For training, validation and testing it uses abstract Blender generated images (size: 540 x 540 x 3) of cubes. |
| Futher it can use an LeNet-5 like architecture (for more see convolutional neural networks on Wikipedia). |
| The whole can be used as a template for other CNN projects. Feel free. |
|
|
| ## The data |
| The github repo https://github.com/osick/collision comes with this large data file set with thousand of generated images of two cubes in three different categories: |
|
|
| * collision: The two cubes toch each other but don't contain themself |
| * in: One cube contains the other |
| * out: The cubes are separated |
|
|
| All gratitude goes to my son, who generated all the images with blender and python scripting. |
|
|
| ## More Info |
| For more infos see https://github.com/osick/collision |
