Simple metrics to evaluate the concealment of an object: SAMI

• Main Authors: Gosseaume, Julien, Kpalma, Kidiyo, Ronsin, Joseph
• Other Authors: Institut d'Électronique et des Technologies du numéRique (IETR), Université de Nantes (UN)-Université de Rennes (UR)-Institut National des Sciences Appliquées - Rennes (INSA Rennes), Institut National des Sciences Appliquées (INSA)-Institut National des Sciences Appliquées (INSA)-CentraleSupélec-Centre National de la Recherche Scientifique (CNRS), Prof. Vaclav Skala
• Format: Conference Object
• Language: English
• Published: HAL CCSD 2014
• Subjects: object segmentation; saliency map; Human Visual System; concealment metrics; Concealment evaluation; object identification; [SPI.SIGNAL]Engineering Sciences [physics]/Signal and Image processing; sami
• Online Access: https://hal.science/hal-01116074; https://hal.science/hal-01116074/document; https://hal.science/hal-01116074/file/SAMI_28042014.pdf

International audience We propose original metrics for estimation of detection and identification of an object in an image: SAMI. SAMI (SAliency based Metrics of Identification) gives a detection score, called D_score, and an identification score, called I_score, for the detection evaluation and the identification evaluation, respectively, for a Region Of Interest (ROI), basically the footprint area of the object. The contribution of this paper is important since SAMI is basically a simple easy-to-implement heuristic method based on existing image processing techniques and some intuition-based postulates. SAMI has initially been conceived to estimate the performance of SCOTT, a “Synthesis COncealment Two-level Texture” algorithm. However, a direct derived application of such metrics could be the evaluation of saliency algorithms for object segmentation: the best segmentation would be the one with the highest SAMI D_score for a given object. Another possible application could be the use of SAMI inside a saliency algorithm, to compute a dense modified saliency map, in which each pixel has the SAMI D_score corresponding to its neigborhood (used as ROI). Such a resulting map would be more robust to saliency noise from small spots.