Home People Research Publications Demos
News Jobs Prospective
About Internal

Top-down Discriminant Saliency and Object Detection

This page presents a collection of results obtained with the proposed top-down discriminant saliency detector (DSD). In all experiments, the detector is trained and tested with unsegmented images containing cluttered backgrounds.

Saliency examples from PASCAL 2006 database

The figure below shows some example saliency maps produced by DSD for various images (and objects classes) from the PASCAL 2006 database. All scenes contain instances from two object classes, e.g. ``person'' and ``car'', or ``motorbike'' and ''car''. In each case, a bounding box is drawn around the object of interest. The saliency map for the detection of that object is shown on the right. It is clear that DSD successfully switches between the two objects, highlighting the one of interest and suppressing all others. This ability is a significant advantage of top-down saliency over bottom-up interest point detection.

Comparison of object localization accuracy on PASCAL 2006 database

In this experiment, the DSD is used to implement a focus-of-attention mechanism that prunes bottom-up (BU) interest points which are not important for the detection of the objects of interest. The object localization accuracy of the selected points (DSD-BU) is measured with precision-recall (PR) curves. The performance of DSD-BU is compared to those of other top-down pruning strategies:

    Discriminant Visual Words (DVW) - clusters of BU points, measuring discriminability by posterior probability of interest points [Bouveyron, et al., ICVGIP 2006; Chum & Zisserman, CVPR 2007].
    linear SVM (LSVM) - clusters of BU points, measuring discriminability by the weight assigned to visual words by a linear SVM classifier [Jurie & Triggs ICCV 2005].
    probabilistic Latent Semantic Analysis (pLSA) - clusters of BU points, based on probabilistic topic discovery for each object class [Sivic et al. ICCV 2005].
In all cases, discriminant saliency achieves substantially better performance.

Comparison of salient locations detected on PASCAL 2006

The figure below shows some examples of the BU interest points selected (at 40% recall rate) by DSD and other methods. From top to bottom: original images (objects marked by their bounding boxes), interest points selected by DSD, DVW, LSVM, and pLSA. Each circle in the image represents the location and size of a salient point. The white color indicates the points which fall inside the segmentation ground truth (the bounding box marked on the original image), while black indicates the opposite.

More detailed DSD results, for each of the object classes in PASCAL 2006, are available here:
bicycle, bus, car, cat, dog, horse, motorbike, person.

Examples from the Brodatz texture database

The figure below shows examples of saliency maps for various textures. These examples show that discriminant saliency can 1) ignore highly textured backgrounds in favor of more salient foreground objects, and 2) detect as salient a wide variety of shapes, and contours of different crispness and scale, or even texture gradients.