Pointers:

• Paper (arxiv preprint) • Dataset • Experimental data

SUIM Dataset

• For semantic segmentation of natural underwater images
• 1525 annotated images for training/validation and 110 samples for testing
• BW: Background/waterbody • HD: human divers • PF: Aquatic plants and sea-grass • WR: Wrecks/ruins
• RO: Robots/instruments • RI: Reefs/invertebrates • FV: Fish and vertebrates • SR: Sea-floor/rocks

SUIM-Net Model

• A fully-convolutional encoder-decoder network: embodies residual learning and mirrored skip connection
• Offers competitive semantic segmentation performance at a fast rate (28.65 FPS on a 1080 GPU)
• Detailed architecture is in model.py; associated train/test scripts are also provided
• The get_f1_iou.py script is used for performance evaluation

Benchmark Evaluation

• Performance analysis for semantic segmentation and saliency prediction
• SOTA models in comparison: • FCN • UNet • SegNet • PSPNet • DeepLab-v3
• Metrics: • region similarity (F score) and • contour accuracy (mIOU)
• Further analysis and implementation details are provided in the paper

Acknowledgements

• https://github.com/qubvel/segmentation_models
• https://github.com/divamgupta/image-segmentation-keras
• https://github.com/floodsung/Deep-Learning-Papers-Reading-Roadmap
• https://github.com/zhixuhao/unet
• https://github.com/aurora95/Keras-FCN
• https://github.com/MLearing/Keras-Deeplab-v3-plus/
• https://github.com/wenguanwang/ASNet