DeepFish: A Realistic Fish-Habitat Dataset

Overview

Fish in the conditions they actually live in

Most underwater datasets are filmed in clear water with a few, well-lit fish. DeepFish comes from remote coastal habitats of tropical Australia: turbid mangroves, seagrass beds and reefs, where fish are small, partly hidden and often absent.

Cameras mounted on metal frames were lowered from a vessel to the seabed and left to record the natural fish community while the vessel kept 100 m away. Depth and GPS position were logged for each deployment. Filming was done in daylight during periods of relatively low turbidity, in full HD.

The footage originally came with classification labels only. We added point-level and segmentation annotations, making DeepFish a benchmark for four tasks.

Tasks

One dataset, four benchmarks

Task	Annotation	Images
Classification	Whether the frame contains a fish	39,766
Counting	Number of fish in the frame, from point labels	3,200
Localization	A point on every fish	3,200
Segmentation	Pixel-level fish masks	620

Diagram: an image passes through a ResNet-50 backbone, then a feed-forward branch for classification and counting and an upsampling branch for localization and segmentation. — Baseline architecture: a shared ResNet-50 backbone with one head per task.

Three panels of the same reef frame: ground-truth points on two fish, the model's two predicted blobs, and the heatmap. — Localization: ground-truth points, predicted fish, and the model's heatmap.

A reef frame with two fish, shown twice with their segmentation masks overlaid. — Segmentation: pixel-level masks for each fish.

Counting and localization subset.

Segmentation subset.

Habitats

Twenty habitats, from reef to mangrove

Frames per habitat in the classification set, split into frames with and without fish. More than half of all frames contain no fish, as in real survey footage.

Habitat	Frames	With fish	Distribution
Total	39,766	17,409

With fishNo fish

Download

Get the data

The dataset is a single 7.1 GB archive. The Hugging Face mirror is fastest from anywhere in the world, and interrupted downloads resume when you rerun the command.

pip install -U huggingface_hub
hf download Alzayats/DeepFish DeepFish.tar DeepFish.tar.sha256 --repo-type dataset --local-dir .
sha256sum -c DeepFish.tar.sha256
tar xf DeepFish.tar

Mirror: huggingface.co/datasets/Alzayats/DeepFish
Original: data.qld.edu.au (James Cook University data portal)
SHA-256: 8acabb8a314fa8eb45fa9cc6829c84bdc127b1f8f605f2e152304e98b6dbb46c
Layout: DeepFish/Classification, Localization, Segmentation

Training and evaluation code for all four tasks is on GitHub.

Citation

Cite DeepFish

If you use the dataset, please cite the paper:

@article{saleh2020realistic,
  title   = {A realistic fish-habitat dataset to evaluate algorithms for underwater visual analysis},
  author  = {Saleh, Alzayat and Laradji, Issam H and Konovalov, Dmitry A and Bradley, Michael and Vazquez, David and Sheaves, Marcus},
  journal = {Scientific Reports},
  volume  = {10},
  number  = {1},
  pages   = {14671},
  year    = {2020},
  publisher = {Nature Publishing Group UK London},
  doi     = {10.1038/s41598-020-71639-x}
}