This paper discusses an algorithm for segmenting the salient regions of a scene using both 3-D laser scan and imagery information. Our problem statement is shown in Fig.1. Given an image and its extrinsically calibrated 3-D scan data, our algorithm delineates top k-most salient regions in
a bottom-up manner without any high-level priors, models, or learning.

Our definition of saliency in a 3-D point cloud is based on a well-known perceptual organization theory - Gestalt laws of grouping. The detection of salient clusters is formulated into finding the set of clusters that best fit some pre-defined families of probability density functions. (Here we use two pdfs, which are Gaussian and Uniform.) As a fundamental tool for this approach, we use the Robust Information-theoretic Clustering (RIC) method. After clustering, we calculate saliency value of each cluster by simply summing four different saliency features, which capture local, regional, and global properties of 3-D data and RGB color information.

Our matching method consists of two steps. First, we find correspondence candidates between linear fragments. Second, a spectral matching algorithm is used to find the subset of correspondences which is the most consistent. Both matching methods are learnt by using logistic classifiers.

Fig.2 shows some experimental results using the dataset taken in parking lots around CMU. The second column represents the results of RIC clustering. The top four most salient regions in the 3-D data and the imagery data are shown in the third and fourth columns, respectively.

(a) Input: an image and its corresponding 3-D scan data (b) Information-theoretic optimal clustering (c) Output: segmentation of top-k most salient regions Figure 1. Problem Statement.

Figure 2. More examples of the detection of salient regions.

Publication