Semantic segmentation using regions and parts

Abstract

We address the problem of segmenting and recognizing objects in real world images, focusing on challenging articulated categories such as humans and other animals. For this purpose, we propose a novel design for region-based object detectors that integrates efficiently top-down information from scanning-windows part models and global appearance cues. Our detectors produce class-specific scores for bottom-up regions, and then aggregate the votes of multiple overlapping candidates through pixel classification. We evaluate our approach on the PASCAL segmentation challenge, and report competitive performance with respect to current leading techniques. On VOC2010, our method obtains the best results in 6/20 categories and the highest performance on articulated objects.

A hybrid framework combining background subtraction and deep neural networks for rapid person detection

“…R-CNN (Regions with CNN features) has been also used for person detection. It reaches 53.9% accuracy of human classification accuracy in the VOC2011 dataset [65], while other region-based methods such as 'Regions and Parts' [66] deliver a slightly lower accuracy. MixedPeds algorithm [67] is a quite original approach.…”

Multi-component Models for Object Detection

“…Therefore, a bounding box selection scheme is necessary to prune out windows in an early stage that do not contain any object. In this paper, we start by applying the segmentation algorithm of [3] which produces a pool of overlaid segments over scales for an input image. Since the algorithm uses gPb contour signals as input which recovers almost full recall of object boundaries at a low threshold, the output segments encode the sizes and shapes of objects in the input image quite precisely.…”

Measures and Meta-Measures for the Supervised Evaluation of Image Segmentation

“…Motivated by the fact that image segmentation is increasingly being used as a preliminary step for object detection [18,1], we propose to assess segmentation under this perspective, that is, we interpret regions in a partition as potential object candidates, and classify them as correct or not. Similarly, we interpret regions in an oversegmentation as parts of objects, if merged together can form an object of the ground truth (inspired by [12] in range image segmentation evaluation).…”

“…In this scenario, bottom-up segmentation methods often play an important role in the proposed algorithms [1,5], and thus improving segmentation techniques would entail improvements towards better semantic segmentation [18]. In such a challenge, providing benchmarks that help researchers understand the weak and strong points of their algorithms is of paramount importance.…”