Adds a maintenance guide page with information on maintaining vision

src/book/03-guides/01-main-codebase/02-maintaining-subsystems.mdx

@@ -0,0 +1,134 @@
+---
+section: Guides
+chapter: Main Codebase
+title: Maintaining Subsystems
+description: How to maintain subsystems within the main codebase.
+slug: /guides/main/maintaining-subsystems
+---
+
+This page details how to maintain various subsystems within the main codebase.
+
+# Vision
+
+Before vision can work on the robot, odometry and kinematics must work on the robot. If these are inaccurate, then the green horizon will be inaccurate.
+
+## Dataset Generation
+
+Synthetic and semi-synthetic training data for vision can be generated using NUpbr and NUgan. Pre-generated datasets for training the Visual Mesh can be found on the Network-Attached Storage (NAS) in the lab.
+
+### NUpbr
+
+NUpbr is a Physically Based Rendering tool created in Blender. It creates semi-synthetic images with corresponding segmentation images for training.
+
+NUpbr requires Blender **2.79**. Download the appropriate file from the [Blender 2.79 Stable Release](https://download.blender.org/release/Blender2.79/).
+
+Clone the [NUpbr repository](https://github.com/NUbots/NUpbr) using
+
+```sh
+git clone https://github.com/NUbots/NUpbr.git
+```
+
+Navigate to the NUpbr directory and run
+
+```sh
+blender --python pbr.py
+```
+
+Find out more about NUpbr and how to edit the parameters on the [NUpbr NUbook page](/system/tools/nupbr).
+
+### NUgan
+
+The NUbots CycleGAN performs style transfer from the semi-synthetic style to a real image style. It creates more realistic images from the semi-synthetic images from NUpbr. It can be used for style transfer to a specific field for fine-tuning vision.
+
+Clone the [NUgan repository](https://github.com/NUbots/NUgan) using
+
+```sh
+git clone https://github.com/NUbots/NUgan.git
+```
+
+The [NUgan repository](https://github.com/NUbots/NUgan) details how to train, test and generate images.
+
+NUpbr images can be found on the NAS for the A datasets. The B dataset can be obtained by taking pictures, or collecting images from [Bit-Bots ImageTagger](https://imagetagger.bit-bots.de/).
+
+### Setting Up The Data
+
+The [Visual Mesh repository](https://github.com/Fastcode/VisualMesh) contains information on how to set up the datasets. It contains [a script](https://github.com/Fastcode/VisualMesh/blob/master/training/build_dataset.py) that takes data and converts it into the tfrecord format that the Visual Mesh expects. The script requires a folder of raw images, segmentation masks, and metadata. NUpbr will provide all these as output, and such data can be found on the NAS. If using output from NUgan, use the corresponding NUpbr segmentation masks and metadata, and add in the generated NUgan images in the image folder.
+
+## The Visual Mesh
+
+### Training
+
+Ensure you have followed the above steps to get a dataset to train on.
+
+As specified in the [Visual Mesh repository README](https://github.com/Fastcode/VisualMesh), install the dependencies from `requirements.txt`.
+
+```sh
+pip3 install -r requirements
+```
+
+Using [`cmake`](https://cmake.org/) and a C++ compiler with C++11 or later, build the custom TensorFlow op.
+
+```sh
+mkdir build
+cd build
+cmake ..
+make
+```
+
+A yaml file must be created with options for training. An example of this is in [the example_net.yaml file](https://github.com/Fastcode/VisualMesh/blob/master/example_net.yaml).
+
+Execute the training by running
+
+```sh
+./mesh.py train [config.yaml] [output_dir]
+```
+
+Where `[config.yaml]` is the options yaml file as specified above, and the output directory is the name of the directory you want the progress and results to be saved to. Progress can be monitored with TensorBoard.
+
+### Testing
+
+Testing the network is executed similarly to training. Run
+
+```sh
+./mesh.py test [config.yaml] [output_dir]
+```
+
+Where `[config.yaml]` and `[output_dir]` are the same as for training. The output will be the average precision for each class, as well as the precision/recall curves and the final mean average precision.
+
+### Exporting Configuration
+
+The weights and biases will be output into a yaml file like the one found at [model.yaml on the Visual Mesh repository](https://github.com/Fastcode/VisualMesh/blob/master/example/model.yaml) or [VisualMesh.yaml on the NUbots repository](https://github.com/NUbots/NUbots/blob/master/module/vision/VisualMesh/data/config/VisualMesh.yaml).
+
+Replace the NUbots repository `VisualMesh.yaml` configuration with the configuration you just obtained from training the Visual Mesh. Ensure you create a new branch so you can make a PR and get any changes merged in. For further information about Git, view the [Git Guide](/guides/general/git).
+
+## Camera Calibration
+
+If the cameras are not calibrated correctly, the vision system will not work well. Information on calibration is available on [the input page](/system/subsystems/input).
+
+An automatic camera calibration tool is available to use in the NUbots repository that can be used to calibrate the cameras. A guide on how to calibrate the cameras will be available on NUbook soon.
+
+## Testing
+
+Now that the Visual Mesh has been updated in the NUbots repository, you should test it before merging.
+
+Build the code, ensuring `VisualMesh.role` is on in `./b configure -i`. Go to the [Getting Started page](/guides/main/getting-started) to find out more on how to build the code and install onto a robot. Ensure the new configuration file has been installed. Check out the [Build System page](/system/foundations/build-system) to find out more about options when installing onto the robot.
+
+When your new Visual Mesh is installed onto the robot, connect to the robot and ensure NUsight is on.
+
+```sh
+nano config/NUsight.yaml
+```
+
+Turn `vision object` and `compressed images` on. Run NUsight using `yarn prod` and navigate to the NUsight page in your browser. More on NUsight can be found on [the NUsight page](/system/tools/nusight).
+
+Run the visualmesh role
+
+```sh
+./visualmesh
+```
+
+Wait for the cameras to load and then watch the Vision tab in NUsight. To determine if the output is correct, consult the [vision page](/system/subsystems/vision) for images of the expected output.
+
+If you would like to see the visual mesh output in NUsight, you will need to log the data and run it back in NUsight using DataPlayback, since the data is too large to send over a network. Add into the `visualmesh.role` file `support::logging::DataLogging`. Build to the robot and rerun the visual mesh binary with `./visualmesh`. An NBS file will be created on the robot in `logs/visualmesh/`. Copy this across the network to the computer you built on using `scp <robot_address>:logs/visualmesh/<file_name> ./recordings` on the computer in the NUbots folder. Read [the Logging page](/system/subsystems/logging) to find out how to play an NBS file back.
+
+If the output has improved, create a PR and get the changes merged in to the GitHub repository. For further information about Git, view the [Git Guide](/guides/general/git).