Vision and Image Processing Lab, Indian Institute of Technology Bombay, India Avik Hati, Subhasis Chaudhuri, Rajbabu Velmurugan [email protected], [email protected], [email protected] Saliency Detection based on a Sparsity of Segments Model Introduction • Saliency: measure of importance of objects in images • Salient object captures attention distinctive in features • Applications Object segmentation Video summarization Content based image compression Image and video quality assessment • Only the important parts are processed • Reduction in complexity Typical Methods • Local approach (center-surround window) Pixel based [Harel et al., Itti et al.] Boundary of large smooth salient object is detected • Global approach (global comparison) Superpixel based [Perazzi et al.] Patch based [Margolin et al., Yan et al.] Salient objects are not uniformly highlighted • Frequency domain [Guo et al., Hou et al.] Problem Definition • Object based approach • Obtain a saliency map of an image such that pixels within an object to have equal saliency salient objects to be extracted completely retain exact boundary of the salient object number of object segmentation to be small Texture Removal Gaussian Mixture Model • EM algorithm assuming GMM • Minimize to find optimum number of Gaussians, K : Gaussian pmfs : mixing proportions • Number of segments is K Input image Histogram Texture-removed image Histogram TV : Total Variation : Weights : Texture-removed piecewise constant image : Regularization parameter i g k e i () i f x 255 0 1 () () arg min k k i i x i k k e hist x f x K e Segmentation • Obtain thresholds by minimizing probability of misclassification , , Saliency Computation • Color saliency : area of region • Spatial saliency : spatial std. dev. of region Salient Object Extraction • Retain exact object boundaries • Image Matting [Levin et al.] • Eroded saliency output as foreground scribble • Dilated saliency output as background scribble Saliency Results Challenges • Choice of regularizer constant Large value removes small objects • Texture-removed image may not always facilitate meaningful segmentation Over smoothing Combination of Y, Cb, Cr components Proposed saliency maps in (e) are complete and without textures, holes, unlike (b-d). Comparison of precision, recall and F-measure of SF, RC, CA with the proposed method (Our) on the MSRA dataset. (see paper for details) i Texture-removed image Segmented regions (K=3) GMM (K=3) 1 0 1 () () i K i K e i i i i P f x dx f x dx 0 0 255 K i Ar i r Segmentation result Color saliency output (thresholded) Spatial saliency Input image-2 Segmentation Color saliency 1,2,..., () 1 max j s j j K s j j Input image Saliency output Salient object Input image (a) Cheng et al. (b) Perazzi et al. (d) Proposed (e) Goferman et al. (c) Multiple salient objects Salient region that can be overlooked by presence of known objects Very small salient objects 2 () , , min i i TV F g i Y Cb Cr g I g 1 () () ( , ) K c i c j i i s j rd r r Precision-Recall Salient object: Butterfly Salient object: Green apple j r