---
license: cc-by-nc-sa-4.0
task_categories:
- image-classification
tags:
- raw
pretty_name: Raw Instinct - A rice grain dataset for RAW image classification
size_categories:
- 1K<n<10K
language:
- en
---

This dataset is from the paper "[Raw Instinct: Trust Your Classifiers and Skip the Conversion](https://doi.org/10.1109/PRAI59366.2023.10332021)" (also on arXiV: https://arxiv.org/abs/2403.14439)

It is also available on Kaggle: https://www.kaggle.com/datasets/mathiasviborg/raw-instinct

This dataset is meant to compare RAW and RGB inputs to deep learning classifiers.
All images are of grains of different rice, and one class has its own folder with training, validation and test data.
RAW files come in 40x40 16-bit .npy files, RGB files come in 40x40 in both 8- and 16-bit .png files.

![image/jpeg](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F13804035%2F3432d6614820c5e1f27a4e8143ee44ac%2Fimage.png?generation=1677845467008506&alt=media)

The 5 classes in the data capture (from left): arborio, basmati, brown, jasmine and parboiled. The dataset consists of 70% training, 20% validation and 10% testing. Distribution of classes are shown in the table below.

![image/jpeg](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F13804035%2F345f719f34954b5af49b64f1a12af4d8%2FSample_distribution.png?generation=1677845296625597&alt=media)

The camera used for the data capture is the 151 MP Phase One XF IQ4 camera, using a Schneider Kreuznach LS 80 mm f/2.8 lens. In order to attain small images (40x40) using the given sensor, the camera was placed 550 cm above the capture surface, where 22,887 samples were captured from a total of 50 high resolution images (457.74 samples from each image on average), stored as IIQ 16-L RAW files. The files were accessed using the Phase One Image SDK and converted to both 8- and 16-bit (per color channel) lossless PNG images for RGB comparison. In this way, three datasets were generated consisting of the same data, one represented in the RAW image format (IIQ) and two in the RGB image format (8- and 16-bit PNG).

![image/jpeg](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F13804035%2F9d5b5953d246a682351232bdf6863bfc%2FCamera_capture-cropped.png?generation=1677845371415847&alt=media)

If you use this dataset in your work, please cite:
```
@INPROCEEDINGS{raw_instinct_2023,
  author={Kantas, Christos and Antoniussen, Bjørk and Andersen, Mathias V. and Munksø, Rasmus and Kotnala, Shobhit and Jensen, Simon B. and Møgelmose, Andreas and Nørgaard, Lau and Moeslund, Thomas B.},
  booktitle={2023 IEEE 6th International Conference on Pattern Recognition and Artificial Intelligence (PRAI)}, 
  title={Raw Instinct: Trust Your Classifiers and Skip the Conversion}, 
  year={2023},
  volume={},
  number={},
  pages={456-460},
  doi={10.1109/PRAI59366.2023.10332021}}
```