Deep learning pipeline for grid classification, segmentation, detection and related problems

Adrien Chan-Hon-Tong (1)

1 ONERA - The French Aerospace Lab [Palaiseau]

We focus on a cluster of computer vision problems containing segmentation, detection and related variants. We benchmark three deep learning pipelines, each trained following a variante of segmentation, on five problems from the cluster. This benchmark is interesting to anyone looking for a deep learning pipeline to deal with multi tasks and/or who has not a clear idea of metric underlying the desired task. In our current results, grid segmentation is the most robust training strategy despite being both the simplest and fastest strategies. Such result may interest the computer vision community if confirmed on larger experiments.

Domaines: Traitement des images [eess.IV]

Adrien Chan-Hon-Tong. Deep learning pipeline for grid classification, segmentation, detection and related problems. 2017. ⟨hal-01412086v4⟩