This work presents a strategy to train a dense semantic segmentation model when there are only a few sparse training labels. The proposed method to augment this sparse labels provides comparable results to training using densely labeled images. We have also worked on building generic encoders on large coral datasets to facilitate more accurate models on new environments.

Authors and collaborators:

Iñigo Alonso, A. Cambra, A. Muñoz, Ana C Murillo

M. Yuval, G. Eval, T. Treibitz

Related results and publications:

Repeatable semantic reef-mapping through photogrammetry and label-augmentation

M Yuval, I Alonso, G Eyal, D Tchernov, Y Loya, AC Murillo, T Treibitz. Remote Sensing 13 (4), 659. 2021.
Paper

CoralSeg: Learning Coral Segmentation from Sparse Annotations.

I. Alonso, M. Yuval, G. Eyal, T. Treibitz, A. C. Murillo. Journal of Field Robotics, 2019, vol. 36, no 8, p. 1456-1477.
Paper (pre-print) | Paper (JFR) | Code | Generic encoder model - TensorFlow| EilatMixx data | Eilat data | Mosaics UCSD data

Semantic Segmentation from Sparse Labeling using Multi-Level Superpixels. 

I. Alonso, A. C. Murillo. IEEE/RSJ Int. Conference on Intelligent Robots and Systems (IROS) 2018.  Paper | Code 

Coral-Segmentation: Training Dense Labeling Models with Sparse Ground Truth.

I. Alonso, A. Cambra, A. Muñoz, T. Treibitz, A. C. Murillo. VWM, ICCV Workshops 2017. Paper | Video

How to adapt existing CNN models for new semantic segmentation tasks?

A. B. Cambra, A. Muñoz, A. C. Murillo. Robot 2017: Third Iberian Robotics Conference. Sevilla, Spain. 2017. Paper

Funding:

The authors would like to thank NVIDIA Corporation for the donation of a Titan Xp GPU used in this work. 

This research has been partially funded by the Spanish Government project PGC2018-098817-A-I00 (MCIU/AEI/10.13039/501100011033/FEDER,UE), UZCUD2017-TEC-06 and Aragón regional government (Grupo DGA T45 17R/FSE, DGA_FSE T45_20R).