Skip to content

A Benchmark for Semantic Segmentation of Waterbody Images

Notifications You must be signed in to change notification settings

rtmccune/atlantis

 
 

Repository files navigation

ATLANTIS - ArTificiaL And Natural waTer-bodIes dataSet

This is the respository for the ATLANTIS. All waterbody labels are comprehensively described in ATLANTIS Wiki. This dataset developed in the iWERS lab at the University of South Carolina.

Figure 1. ATLANTIS - ArTificiaL And Natural waTer-bodIes dataSet.

Overview

ATLANTIS is a benchmark for semantic segmentation of waterbody images. For the first time, this dataset covers a wide range of natural waterbodies such as sea, lake, river and man-made (artificial) water-related sturcures such as dam, reservoir, canal, and pier. ATLANTIS includes 5,195 pixel-wise annotated images split to 3,364 training, 535 validation, and 1,296 testing images. In addition to 35 waterbodies, this dataset covers 21 general labels such as person, car, road and building.

AQUANet

In addition to waterbodies dataset, and in order to tackle the inherent challenges in the segmentation of waterbodies, we (iWERS in collaboration with Computer Vision Lab) developed a CNN-based semantic segmentation network which takes advantage of two different paths to process the aquatic and non-aquatic regions, separately. Each path includes low-level feature and cross-path modulation, to adjust features for better representation. The results show that AQUANet outperforms other state-of-the-art semantic-segmentation networks on ATLANTIS, and the ablation studies justify the effectiveness of the proposed components. The weights of trained models in this study are publicly availble in Google Drive.

Figure 2. The network architecture of proposed AQUANet.

Dataset Description

The ATLANTIS dataset is designed and developed with the goal of capturing a wide-range of water-related objects, either those exist in natural environment or the infrastructure and man-made (artificial) water systems. In this dataset, labels were first selected based on the most frequent objects, used in water-related studies or can be found in real-world scenes. Aside from the background objects, total of 56 labels, including 17 artificial, 18 natural water-bodies, and 21 general labels, are selected. These general labels are considered for providing contextual information that most likely can be found in water-related scenes. After finalizing the selection of waterbody labels, a comprehensive investigation on each individual label was performed by annotators to make sure all the labels are vivid examples of those objects in real-world. Moreover, sometimes some of the water-related labels, e.g., levee, embankment, and floodbank, have been used interchangeably in water resources field; thus, those labels are either merged into a unique group or are removed from the dataset to prevent an individual object receives different labels.

In order to gather a corpus of images, we have used Flickr API to query and collect 800 "medium-sized" unique images for each label based on seven commonly used "Creative Commons" and "United States Government Work" licenses. Downloaded images were then filtered by a two-stage hierarchical procedure. In the first stage, each annotator was assigned to review a specific list of labels and remove irrelevant images based on that specific list of labels. In the second stage, several meetings were held between the entire annotation team and the project coordinator to finalize the images which appropriately represent each of 56 labels. Finally, images were annotated by annotators who have solid water resources engineering background as well as experience working with the CVAT, which is a free, open source, and web-based image/video annotation tool. If you wish to contribute to this project or you want to develop a semantic segmentation dataset, please check ATLANTIS DEVELOPMENT KIT. If you are curious how images are annotated in this project, please watch the follwoing tutorial videos:

(Watch Video 1) (Watch Video 2) (Watch Video 3)

Defining Waterbody Classes

In order to decide what waterbodies should be included in ATLANTIS, an initial list of natural waterbodies and man-made hydraulic structures was prepared. Then, with the objects and labels with common functionalities were merged into a unique label. For example, the pier, dock, and harbor labels, were all labeled into pier, levee, embankment, and floodbank were labeled as levee. Further, for the natural waterbodies, labels were merged based on their visual features. For example, ocean, sea, gulf, and lagoon were grouped and labeled as sea. In addition to introducing a wide range of waterbodies, the ATLANTIS dataset aims to provide contextual information by identifying about auxiliary objects which most likely can be found in water-related scenes. During defining the labels, a list of general objects which can give clues about existence of water-related objects in a scene were identified and considered for annotation. For example, it is expected to see common urban features, such as building and roads, when one wants to identify the different between a river and man-made canal. It is expected that if urban features are found, the waterbody is a man-made canal and not a river. In other words, canals are located where a bunch of pixels are already labeled as "building", "road" or "sidewalk".

Figure 3 demonstrates the spatial distributions of the most frequent co-occurred labels with respect to "river" and "canal" waterbody label. In the other words, Figure 3 indicates the most frequent label at each pixel of the ground truth segmentation tensor which is comprised of all corresponding ground-truth segmentation maps for each label. This figure explicitly approves that sorrunding settings which commonly exist in a "river" scene is totally different from those of "canal."

Figure 3. The most frequent label at each pixel of (a) river, (b) canal.

Another example is when we deal with complex shapes and labels, such as “swamp”. It is very difficult to visually identify the difference between “marsh” and “swamp” even for water resources expert.

Figure 4. (a) Swamps are dominated by trees, (b) while marshes are often dominated by grasses.

In this case, auxiliary labels such as “cypress tree” can help since this kind of tree inhabits exclusively in swamps and has a rather unique shape. Also, herbaceous plants that generally grow in other type of wetlands were considered as part of the “marsh”, and aquatic parts for both cases are annotated as “wetland” which is a general form of such waterbodies. By annotating different types of vegetation as “cypress tree” and “marsh” along with “wetland”, it is expected that models consider co-existence of these labels in a scene to distinguish the differences between similar objects.

Dataset Statistics

Figure 5 shows the frequency distribution of the number of images for waterbody labels. Labels are ranked based on pixel frequencies.

Figure 5. Frequency distribution of the number of images assigned to each waterbody label.

Figure 6 shows the frequency distribution of the number of pixels for all 56 ATLANTIS labels plus background (percentage).

Figure 6. Percentage of pixels frequency distribution for all 56 ATLANTIS labels plus background.

Such a long-tailed distribution is common for semantic segmentation datasets even if the number of images that contain specific label are pre-controlled. Such frequency distribution for pixels would be inevitable for objects existing in real-world. Taking "water tower" as an example, despite having 219 images, the number of pixels are less than many other labels in the dataset. In total, only 4.89% of pixels are unlabeled, and 34.17% and 60.94% of pixels belong to waterbodies (natural and artificial) and general labels, respectively.

Figure 7. The "R-squared" of the regression between the number of images and pixels is relatively low.

Reference

If you use this data, please cite the following paper which can be downloaded through this link:

@article{erfani2022atlantis,
  title={ATLANTIS: A benchmark for semantic segmentation of waterbody images},
  author={Erfani, Seyed Mohammad Hassan and Wu, Zhenyao and Wu, Xinyi and Wang, Song and Goharian, Erfan},
  journal={Environmental Modelling \& Software},
  pages={105333},
  year={2022},
  publisher={Elsevier}
}

About

A Benchmark for Semantic Segmentation of Waterbody Images

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 97.0%
  • Shell 3.0%