[Paper Review] ATLANTIS: A Benchmark for Semantic Segmentation of Waterbody Images

Seyed Mohammad Hassan Erfani, Zhenyao Wu+3 more|Scholar Commons (University of South Carolina)|Nov 22, 2021

Flood Risk Assessment and ManagementEnvironmental Science25 references48 citations

TL;DR

ATLANTIS provides the largest-scale pixel-level annotated dataset for waterbody semantic segmentation and introduces AQUANet, a dual-path network tailored for aquatic and non-aquatic regions, achieving state-of-the-art results on ATLANTIS.

ABSTRACT

Vision-based semantic segmentation of waterbodies and nearby related objects provides important information for managing water resources and handling flooding emergency. However, the lack of large-scale labeled training and testing datasets for water-related categories prevents researchers from studying water-related issues in the computer vision field. To tackle this problem, we present ATLANTIS, a new benchmark for semantic segmentation of waterbodies and related objects. ATLANTIS consists of 5,195 images of waterbodies, as well as high quality pixel-level manual annotations of 56 classes of objects, including 17 classes of man-made objects, 18 classes of natural objects and 21 general classes. We analyze ATLANTIS in detail and evaluate several state-of-the-art semantic segmentation networks on our benchmark. In addition, a novel deep neural network, AQUANet, is developed for waterbody semantic segmentation by processing the aquatic and non-aquatic regions in two different paths. AQUANet also incorporates low-level feature modulation and cross-path modulation for enhancing feature representation. Experimental results show that the proposed AQUANet outperforms other state-of-the-art semantic segmentation networks on ATLANTIS. We claim that ATLANTIS is the largest waterbody image dataset for semantic segmentation providing a wide range of water and water-related classes and it will benefit researchers of both computer vision and water resources engineering.

Motivation & Objective

• Motivate semantic segmentation for waterbodies and water-related scenes to support water resource management and flood response.
• Create a large-scale, richly annotated dataset covering natural and artificial waterbodies plus general scene labels.
• Develop AQUANet, a two-path network that processes aquatic and non-aquatic regions separately to improve segmentation of water-related classes.
• Provide texture-focused analysis via the ATLANTIS Texture (ATeX) subset to study texture cues in waterbodies.
• Evaluate state-of-the-art segmentation models on ATLANTIS to establish baselines and demonstrate AQUANet's effectiveness.

Proposed method

• Curate and annotate ATLANTIS with 5,195 images and 56 pixel-level labels (17 artificial, 18 natural, 21 general).
• Propose AQUANet with two parallel branches (aquatic vs. non-aquatic) built on ResNet-101 backbones, including low-level feature modulation and cross-path modulation.
• In each path, apply low-level feature modulation using F_l from ResNet-101 and an ASP-OC style module to produce probability maps.
• Apply cross-path modulation to refine aquatic and non-aquatic probability maps before concatenation and upsampling to the original image size.
• Train with standard cross-entropy loss plus an auxiliary loss on an intermediate ResNet feature, with 1.0 main loss weight and 0.4 auxiliary weight.
• Create ATLANTIS Texture (ATeX) with 12,503 patches labeled by waterbody category for texture-focused analysis.

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

Experimental results

Research questions

• RQ1 Can a large, diverse, pixel-annotated dataset of waterbodies enable training robust semantic segmentation models for water-related scenes?
• RQ2 What is the performance gap between generic semantic segmentation models and a water-specialized architecture on ATLANTIS?
• RQ3 Do two-path architectures and feature modulation strategies improve segmentation of visually similar waterbody classes?
• RQ4 How do low-level texture features and cross-path interactions influence aquatic class delineation?

Key findings

• ATLANTIS comprises 5,195 annotated images across 56 classes, enabling broad coverage of waterbodies and related objects.
• AQUANet consistently outperforms ten state-of-the-art segmentation networks on ATLANTIS across aquatic and non-aquatic categories in quantitative tests.
• Two-path design with aquatic and non-aquatic branches yields notable gains in aquatic region accuracy and IoU (A-acc, A-mIoU).
• Low-level feature modulation and cross-path modulation each contribute to performance gains, with ablations confirming their effectiveness.
• ATeX provides a texture-focused dataset for waterbody texture analysis, enabling texture-based classification study alongside semantic segmentation.

(a)