Object Extraction from Bounding Box Prior with Double Sparse Reconstruction

Abstract

Extracting objects from natural images has long been an active problem in image processing. Despite various attempts, it has not been completely solved up to date. Current state-of-the-art object proposal methods tend to extract a set of object segments from an image, and often these are consequential differences among these results for each image. Another type of methods strive to detect one object into a bounding box where some background parts are often covered. For these two methodologies, we observe: 1) there are generally some regions overlapped among different proposals, which are usually from one object; they could be as object 'segment hypotheses'; 2) pixels outside the detected bounding box could be as 'background hypotheses' as they are with high probability from the background. With them, we formulate the object extraction as a "double" sparse reconstruction problem in terms of the bounding box results. The idea is that object regions should be with small reconstruction errors to segment hypotheses bases, simultaneously, they should have large reconstruction errors to background hypotheses bases. Comprehensive experiments and evaluations on PASCAL VOC object segmentation dataset and GrabCut-50 database demonstrate the superiority of our built method. In particular, we achieve the state-of-theart performance for the object segmentation with bounding box prior on these two benchmark datasets.

Superpixel Sizes using Topology Preserved Regular Superpixel Algorithm and their Impacts on Interactive Segmentation

“…When compared to state-of-the-art interactive segmentation algorithms, 2500 m3 achieved 0.024, which is 0.001 higher than the best reference algorithm in the error rate category, which is a training-based algorithm (see Table III). It did, however, outperforms other training-based algorithms like Label propagation [43] (0.0321), Xia [44] (0.033), SBT with AT [44] (0.033,) RW with AT [45] (0.033), and DEEPGC [38] (0.034). Apart from this, in terms of the F-score (see Table III) and Jaccard index (see Table IV), there is a 4% difference between the best reference algorithm and our algorithm.…”