Overview

In this project we examined the automatic seagrass coverage estimation of the sea bottom. A total of 12682 images of the seabed at different depths along Croatia's Adriatic coast were taken with the help of a diving robot. Of these, 6036 images were manually polygon-annotated by hand and made available to the public as pixel maps. Using this dataset, we tested a superpixel classification of seagrass images. To achieve this, we used several different feature extraction methods as for example CNN-Features that turned out to be the best ones in our experiments.

This project is a joint work between University of Zadar - Croatia and University of Applied Sciences Fulda - Germany

Envolved People

• Gereon Reus, Fulda
• Thomas Möller, Fulda
• Jonas Jäger, Fulda
• Julian Hasenauer, Fulda
• Dr. Stewart T. Schultz, Zadar
• Dr. Claudia Kruschel, Zadar
• Dr. Viviane Wolff, Fulda
• Dr. Klaus Fricke-Neuderth, Fulda

Paper

The full paper Looking for Seagrass: Deep Learning for Visual Coverage Estimation (IEEE - OCEANS 2018 Kobe) is available at: HS Fulda

Dataset

looking-for-seagrass-dataset

Deep Net for feature extraction

http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz

Experiments

Getting started

Dependencies

• Anaconda 3
• CUDA 8 & cuDNN 6
• VirtualEnv (Dependencies located in conda-env.yml)

Checklist

1. Install dependencies
2. Create VirtualEnv
3. Download InceptionNet V3 Graph
4. Download LookingForSeagrass Dataset (and extract)
5. Adapt three paths in experiment scripts
6. Activate Virtual Env
7. Run experiments

Paths in experiments

#Path to InceptionNetV3 ProtoBuf
GRAPH="/path/to/classify_image_graph_def.pb"

# Root Path of LookingForSeagrass Dataset
FOLDER_ROOT="/path/to/datasetroot/dataset"

# Path for storing your results
OUTPUT_PATH="/path/to/output/results"

Citation

Please cite our Paper:

@INPROCEEDINGS{8559302,
author={G. Reus and T. M{\"o}ller and J. J{\"a}ger and S. T. Schultz and C. Kruschel and J. Hasenauer and V. Wolff and K. Fricke-Neuderth},
booktitle={2018 OCEANS - MTS/IEEE Kobe Techno-Oceans (OTO)},
title={Looking for Seagrass: Deep Learning for Visual Coverage Estimation},
year={2018},
volume={},
number={},
pages={1-6},
keywords={},
doi={10.1109/OCEANSKOBE.2018.8559302},
ISSN={},
month={May},}