Using ResNet-18 to correctly identify animal species in images of the Island Conservation dataset, which is freely available, but not featured in any publications yet.
Characteristics of the dataset:
- ca. 120,000 high-resolution images in different sizes
- 123 camera locations from 7 islands in 6 countries
- highly imbalanced, 49 classes overall
- 60% of images are empty without an animal
- ca. 6,000 images with more than one animal
The state-of-the art for identifying the species correctly using ResNet-18 is at 98% accuracy (Tabak et al. (2018) who use a different dataset).
Create a conda environment from the .yml file. Use Code/model_training.py for training + validating. Metrics are reported to local Tensorboard.
This individual project is ongoing as part of my coursework for my Masters degree (Management & Data Science) in the winter term 2020/2021.
The Island Conservation Camera Traps Dataset is hosted by the Labeled Information Library of Alexandria: Biology and Conservation (LILA BC) and is freely available here.
The ResNet-18 implementation and pretrained weights are provided by PyTorch here.
Tabak, M. A., Norouzzadeh, M. S., Wolfson, D. W., Sweeney, S. J., Vercauteren, K. C., Snow, N. P., Halseth, J. M., Di Salvo, P. A., Lewis, J. S., White, M. D., Teton, B., Beasley, J. C., Schlichting, P. E., Boughton, R. K., Wight, B., Newkirk, E. S., Ivan, J. S., Odell, E. A., Brook, R. K., … Miller, R. S. (2018). Machine learning to classify animal species in camera trap images: Applications in ecology. Methods in Ecology and Evolution, 10(4), 585–590. https://doi.org/10.1111/2041-210X.13120