[CV][Descriptor] Ray Feature

1. Authors

Smith et al. proposed Ray feature in [1].

2. Successfully applied area:

- Cell detection

3.Main idea

A Ray feature set contains four kinds of features (Figure 1):

- Distance difference feature $f_i^{[ diff ]}$: provide the capability of scale invariance--> reduce the number of features in feature pool

- Distance feature $f_i^{[ dist ]}$: provide distance contraint on scale

- Orientation feature $f_i^{[ ori ]}$: provide relative gradient angle

- Norm feature $f_i^{[ norm ]}$: provide norm of gradient vector

Figure 1: Four kinds of Ray feature

Four feature templates are created in feature pool. Figure 2 shows that in nuclei, mitochondria and face detection, distance feature weakly contributes to detection while distance difference feature seems good.

- Precomputation like integral image is also proposed to help compute features quickly. According to authors, precomputation process is lower than Haar precomputation but in runtime, their features are faster (Figure 3).

Figure 2: The template contribution for problem of nuclei, mitochondria and face detection

Figure 3: Precomputation and run-time cost for Rays and Haar

4. My conclusion:

- Feature for narrow area.

- Good for deformable object (Figure 4) in simple background and good-result-for-edge-detection image, for example: nuclei and mitochondria in cell environment

- Adaboost/SVM is model for automatically choosing good features in a set of created features from some feature templates.

Figure 4: cell and its shape

[1] K. Smith, A. Carleton, and V. Lepetit. "Fast Ray Features for Learning Irregular Shapes". In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Kyoto, Japan, 2009.