DARMA: Detection as A Reinforced Means of Attention

Matthew Keaton · Ram Zaveri · Meghana Kovur · Cole Henderson · Gianfranco Doretto

Paper | arXiv | Project Page

Challenge

We tackle the challenege of "fine-grained visual classification of plant species in the wild" using the data from the challenge PlantCLEF2015. Here, we reinforce the classification task by reducing the classification to various plant organs and then fusing the final results. We do that by performing object detection onto those organs. We make our curated data and our annotation tool publicly available in the following section.

Curated Data

Please download the data from our shared Google Drive link. We collect data from the challenge PlantCLEF2015, and annotate plant organs using our custom-made annotation tool in the following classes:

• leaf
• fruit
• flower
• bark
• HDL: High Density Leaves

The challenge provides two splits:

• train
• test

The train set provides 1000 species to train on and the test set provides 975 species to evaluate on. Additionally, the data is skewed; therefore, we further scrap the internet for more data which we also make available through our shared Google Drive link. The splits are in the following format:

• train_split.zip : train split
• train_extra_web_images.zip : additional train data through web scrapping
• test_split.zip: test split

Each of them contain data in the following format:

• <split>/species
  • <id>.jpg : image
  • <id>.xml : corresponding metadata
  • <id>_annotations.xml: corresponding annotations

Note: the data from the internet does not have metadata.

@inproceedings{keatonZKHAD21cvprw,
  author = {Keaton, M. R. and Zaveri, R. J. and Kovur, M. and Henderson, C. and Adjeroh, D. A. and Doretto, G.},
  title = {{Fine-Grained Visual Classification of Plant Species In The Wild: Object Detection as A Reinforced Means of Attention}},
  booktitle = {Proceedings of the IEEE CVPR Workshop on Fine-Grained Visual Categorization},
  year = {2021},
}