Coral Reef annotation, localisation and pixel-wise classification using Mask R-CNN and Bag of Tricks

0536765 - ÚTIA 2021 RIV DE eng C - Konferenční příspěvek (zahraniční konf.)
Picek, L. - Říha, A. - Zita, Aleš
Coral Reef annotation, localisation and pixel-wise classification using Mask R-CNN and Bag of Tricks.
CEUR Workshop Proceedings : Volume 2696. Working Notes of CLEF 2020 - Conference and Labs of the Evaluation Forum. Achen: CEUR-WS.org, 2020, č. článku 83. ISSN 1613-0073.
[CLEF 2020. Thessaloniki (GR), 22.09.2020-25.09.2020]
Grant ostatní: GA MŠk(CZ) LO1506
Institucionální podpora: RVO:67985556
Klíčová slova: Deep Learning * Computer Vision * Instance Segmentation
Obor OECD: Robotics and automatic control
http://library.utia.cas.cz/separaty/2020/ZOI/zita-0536765.pdf

This article describes an automatic system for detection, classification and segmentation of individual coral substrates in underwater images. The proposed system achieved the best performances in both tasks of the second edition of the ImageCLEFcoral competition. Specifically, mean average precision with Intersection over Union (IoU) greater then 0.5 (mAP@0.5) of 0.582 in case of Coral reef image annotation and localisation, and mAP@0.5 of 0.678 in Coral reef image pixel-wise parsing. The system is based on Mask R-CNN object detection and instance segmentation framework boosted by advanced training strategies, pseudo-labeling, test-time augmentations, and Accumulated Gradient Normalisation. To support future research, code has been made available at: https://github.com/picekl/ImageCLEF2020-DrawnUI.
Trvalý link: http://hdl.handle.net/11104/0314743