This repository contains the code and neural networks that have been used to gain the results in my Master's thesis paper "Super-Resolution of Bathymetric Data Using Diffusion Models".
This work is also a direct extension of J.Shims et.al work so alot of the code is taken from their amazing work. Here you can find their paper and their repository.
The data for training the diffusion model in this paper has been collected by the AUV called RAN, which was operated by Gothenburg University in 2022. It was equipped with an EM2040, a high resolution MBES and the data is in the form of point clouds. The location where the data was collected is Antarctica, more specifically, the Dotson Ice Shelf in west Antarctica. Unfortunately, I cannot share this dataset so this repository is specifically aimed towards those that have access to it.
Data samples from said dataset can be generated with data_gen.py. This script will generate random slices of the pointcloud and save an npz file that can later be used to generate Signed Distance Functions (SDF)s. This is done by running SDF_creator.py. Be sure to load the correct npz file. If you also want intensity data in your dataset, you need to run data_gen_intensity.py which will use the npz file created by data_gen.py and add the corresponding intensity data. It will then save a new npz file which should then be used in SDF_creator.py.
To run the training the command is:
python3 main.py config/sr32_64/sonar.yaml --gpus 0,1,2,....where if 0 is the first gpu, 1 is the second and so on. You can use as many gpus as you want. sonar.yaml is the configuration file. There are some variables there that are more important than others which are explained below.
Important variables New Loss Model (NLM) (explained in the paper):
- new_loss: There are two places where this variable occurs; under the ddpm and the preprocessor. BOTH has to be set to "yes" for the model to be trained with the modified loss explained in the paper.
- sdf_weight: This variable is set as gamma in the paper and it decides how much the model should focus on having correct distances between two adjacent cells in the outputet SDF
Important variables for models using intensity: in_channel: This is the number of input channels in the network. By default, this is 2. However, we want to add the intensity to another channel so it needs to be set to 3. use_intensity: set this to "yes" to train the model with intensity data
To test the networks performance and compare it against an interpolation technique, you can run the networkvsinterp.py script. You can adjust variables in the script that are marked such as what interpolation technique to compare against. An important thing to note is that the configuration file (sonar.yaml) has to be configured for the model you are testing. For example, the new_loss variables have to be "yes" if you test a NLM and in_channel and useintensity has to be set to "yes" if you're testing a model that has been trained with intensity values.
You can find the pretrained networks that have been used in my paper here. Make sure to put each network folder into the SDF-Diffusion/results/sr32_64 folder.