Automatic water gauge reading is very important for cargo

weighting in ocean transportation. In this process, accurate waterline

extraction is an important yet challenging step. Waterline extraction is

subjected to many environmental interference factors, e.g., bad illumi-

nation, bad weather conditions, ambiguous contours of water stain, etc.

In this paper, we propose a joint multitask based deep model for ac-

curate waterline extraction. The proposed model consists of two main

branches. One branch is used to extract high-level contextual informa-

tion. The other branch rooted in shallow layers is used to extract low-level

detail features. The two branches are later coupled with each other to co-

supervise the estimation of waterline. Our model works well on various

conditions, such as uneven light, serious reections, etc. We also intro-

duce a new benchmark dataset for waterline extraction. This dataset

consists of 360 pictures extracted from 69 videos collected in several ac-

tual ports. Furthermore, su cient experiments show that our model is

e ective on the introduced dataset and outperforms the state-of-the-art

methods.