First Task - Konnsy/REAML2022-hackathon Wiki
In this task of the hackathon we provide positive and negative datasets containing raw images, i.e. the recordings that were not preprocessed. Since you cannot spot particles of interest on a raw image, you have to implement a preprocessing yourself. You may use the approach from the presentation or come up with an own solution for an improved version.
We provide a data loader in pytorch that takes the path to a dataset and a window size and returns blocks of the given window size.
Your main task is to make a binary discrimination between samples with particles and samples without particles as reliable as possible. To score the submitted model we will calculate the share of correctly classified images in new datasets where only the presence or absence of particles is taken into account.
For a mobile sensor system which it is desirable for the sensor for detecting virus particles to keep the computing unit small. For this reason, the approach developed should also be able to run on an embedded device, in this case an Odroid N2+. This has only limited computing capacity - no CUDA-capable graphics card and only 4GB of RAM. In return, its physical dimensions are just as mean in terms of energy consumption.
Data and Code Characteristics
You receive sets of raw images for training and validation. Some containing particles of interest and some containing only imaging artifacts (also called disturbances). Feel free to augment the training data or change the training code as only your final classification method will be scored.
Further explanation and visualization of the data recording process:
- Animation of the PAMONO Sensor recording particle images
- Presentation (PDF) of the data characteristics
In https://tu-dortmund.sciebo.de/s/IV5R5q5trCQnSgj9 (Password giRT%6fe$fre) you can find positive (dataset_pos, dataset_pos_test) and negative (dataset_neg, dataset_neg_test) example images. Extract them to the code folder "HackathonREAML2022_Classification", since this is the given standard folder. Alternatively you can also change these paths as well as the output folder in "data_paths.py".
Restrictions for the optional challenge
Please make sure to use no more than 6GB VRAM and 6GB RAM. In case you submit a model for execution on the Odroid, consider that only 4GB are available together as shared RAM/VRAM. For a more safe evaluation than just relying on the hardware parameters, you can also test your model directly on of the provided odroid devices.
In order to keep the results easier to compare, test system will have the libraries specified in the requirements.txt file installed. Please refrain from relying on other libraries without asking. Contact us if you are not sure if you are allowed to use certain functionalities.
In case of problems and questions
If you encounter problems with the provided code or data, or have questions, you can contact [email protected].