Benchmarking losses for deep learning laxness detection

Adrien Chan-Hon-Tong (1)

1 ONERA - The French Aerospace Lab [Palaiseau]

In object detection, classical rule for accepting a match between a ground truth area and a detection is a 0.5 jacard ratio. But does user care about this ? Especially, does deep network training care about this ? And if yes, may user accept some relaxation of the problem if it can help the training ? In this paper, we add laxness on object detection in remote sensing image by deep learning pipeline. We present several models and relaxation of the classical detection problem. Our preliminary results on different public dataset show that some relaxations are very facilitative to train the pipeline outperforming the same pipeline learned for strict detection.

Domaines: Traitement des images [eess.IV]

Adrien Chan-Hon-Tong. Benchmarking losses for deep learning laxness detection. 2017. ⟨hal-01412086v2⟩