[CV][Detector] SFOP

1. Main points:

Forstner et al. proposed SFOP detector based on [2, 3] and automatic-scale selection of Lindeberg (Idea: improve available detector with automatic-scale selection theory).

Interesting things in their paper are:

+ They defined an property of interpretability for detector evaluation. For this consideration, they show that only Lowe, MSER and SFOP have powerful interpretation. Consideration is against to thought of anything that has hight repeatability is good interest point (Is it really good idea?).

+ SFOP performs light better than some scale-invariance state of the art detectors (Figure 2) and in poorly textured image (Figure 3).

+ SFOP gives more correspondences than other detectors (Figure 2: Row 3, first image).

Figure 1: Is interpretability important?

Figure 2: SFOP is light better scale-invariant detectors on Boat set (first row).

Figure 3: Good result of SFOP on poor textured image

2. My Conclusion:

- Should try SFOP for your application.

- SFOP may be your for image visualization (as well as MSER and Lowe)

References:

[1] Wolfgang Förstner, Timo Dickscheid, Falko Schindler. "Detecting Interpretable and Accurate Scale-Invariant keypoints". ICCV. 2009

[2] L. Parida, D. Geiger, and R. Hummel. Junctions: Detection, Classification and Reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1998.

[3] J. Bigun. A Structure Feature for Some Image Processing Applications Based on Spiral Function. Computer Vision, Graphics and Image Processing, 1990.