Adaptive Superpixel Segmentation Aggregating Local Contour and Texture Features

Project summary:


Superpixel segmentation produces atomic regions of pixels (namely, the superpixels) that are consistent with human perception. Unlike the traditional rigid pixel representation of images, superpixels provide visually meaningful entities, which can further be utilized as inputs for mid- or high-level computer vision tasks. The prominent advantage of using superpixels instead of pixels is the reduction in computational cost for subsequent processing. We propose a novel superpixel segmentation method based on an iterative and adaptive clustering algorithm that embraces color, contour, texture, and spatial features together (ACS). The algorithm adjusts the weights of different features automatically in a content-aware way, so as to fit the requirements of various image instances. More specifically, in each iteration, the weights in the aggregation function are adjusted according to the discriminabilities of features in the current working scenario. This way, the algorithm not only possesses improved robustness but also relieves the burden of setting the parameters manually. Experimental verification shows that the algorithm outperforms existing peer algorithms in terms of commonly used evaluation metrics, while using a low computational cost.

Research highlight:

Compared with previous works, the novelties of ACS lie in: (1) perception consistent: we adopt a color difference measure based on a more perception consistent standard; (2) content adaptability: the aggregating weights of features are automatically adapted according to their discriminabilities on different images; and (3) simplicity and efficiency: it is simple and intuitive to add new features on existing superpixel algorithms, but it is a nontrivial work (always tedious and time-consuming) to set reasonable weights. ACS relieves the burden since the weights are automatically set.

Research results:


1. Quantitative evaluation of different superpixel algorithms. (a) Boundary Recall. (b) Undersegmentation Error. (c) Achievable Segmentation Accuracy. (d) Runtime.


2. Visualization of the maximum segmentation performance using different superpixel algorithms when superpixel number is 400.


To be added.