Add paf22 traffic light detection evaluation docs

doc/06_perception/experiments/traffic-light-detection_evaluation/README.md

@@ -0,0 +1,43 @@
+# Evaluation of the PAF22 Traffic Light Detection
+
+In this experiment, the existing Traffic Light Detection from PAF22 has been tested.
+The goals was to be able to verify, that it is suitable for PAF23.
+
+## Model
+
+The architecture of the model is a Convolutional Neural Network (CNN) and it consists of the following layers:
+
+1. **Convolutional Layer 1**: This layer uses a 2D convolution over an input signal composed of several input planes, with in_channels input channels, 4 output channels, a kernel size of 5, and padding set to 'same'. This means the output size is the same as the input size.
+2. **Batch Normalization Layer**: This layer applies Batch Normalization over a 4D input (a mini-batch of 2D inputs with additional channel dimension) as described in the paper Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift.
+3. **Convolutional Layer 2**: This layer is similar to the first convolutional layer but it takes the output of the first layer (4 channels) as input.
+4. **Max Pooling Layer 1**: This layer uses a 2D max pooling over an input signal composed of several input planes, with a kernel size of (2, 2).
+5. **Convolutional Layer 3**: This layer is similar to the previous convolutional layers but it has a kernel size of 3.
+6. **Max Pooling Layer 2**: This layer is similar to the first max pooling layer.
+7. **Convolutional Layer 4**: This layer is similar to the previous convolutional layers.
+8. **Max Pooling Layer 3**: This layer is similar to the previous max pooling layers.
+9. **Flatten Layer**: This layer flattens the input by removing the spatial dimensions.
+10. **Dropout Layer**: This layer randomly zeroes some of the elements of the input tensor with probability p=0.3 using samples from a Bernoulli distribution.
+11. **Linear Layer**: This layer applies a linear transformation to the incoming data. It has 64 input features and num_classes output features.
+
+## Dataset
+
+The existing dataset of PAF22 consists of 2340 images (combined) of the categories red, yellow, green, backside. There are also 382 validation images (combined).
+
+The data can be accessed through DVC.
+
+## Training
+
+Running the training with `dvc exp run` in the traffic light detection directory, results in a trained model with >99% accuracy & validation.
+
+## Examples
+
+Result | Large | Small |
+|-----------|----------|----------|
+Green | ![Green-Large](assets/green_4.png)  |  ![Green-Small](assets/green_22.jpg)
+Yellow | ![Yellow-Large](assets/yellow_1.png)  |  ![Yellow-Small](assets/yellow_18.jpg)
+Red | ![Red-Large](assets/red_10.png)  |  ![Red-Small](assets/red_20.png)
+Back | ![Back-Large](assets/back_1.png)  |  ![Back-Small](assets/back_14.jpg)
+
+## Verdict
+
+The high accuracy and manual testing of the above example images verified, that the existing PAF22 traffic light detection model can be used for PAF23.