Pointing only for remote sensing images

Adrien Chan-Hon-Tong (1)

1 ONERA - The French Aerospace Lab [Palaiseau]

There are a lot of works aiming to reduce the need of human annotations: self supervised training, weakly supervised training, interactive verification instead of annotation. Here, we show on several public dataset that, in remote sensing context, pointing objects i.e. giving the center of each object can be sufficient for learning a segment before pointing pipeline. Resulting pipeline also reaches honorable score on segmentation.

Domaines: Traitement des images [eess.IV]

Adrien Chan-Hon-Tong. Pointing only for remote sensing images. 2017. ⟨hal-01412086v5⟩