Going into depth: Evaluating 2D and 3D cues for object classification on a new, large-scale object dataset

Abstract

Categorization of objects solely based on shape and appearance is still a largely unresolved issue. With the advent of new sensor technologies, such as consumer-level range sensors, new possibilities for shape processing have become available for a range of new application domains. In the first part of this paper, we introduce a novel, large dataset containing 18 categories of objects found in typical household and office environments-we envision this dataset to be useful in many applications ranging from robotics to computer vision. The second part of the paper presents computational experiments on object categorization with classifiers exploiting both two-dimensional and three-dimensional information. We evaluate categorization performance for both modalities in separate and combined representations and demonstrate the advantages of using range data for object and shape processing skills

Robust Instance Recognition in Presence of Occlusion and Clutter

“…Unlike other existing datasets [15,4,22] this dataset has high clutter and pose variations, thus closely approximating real world conditions. ACCV3D contains instances of 15 objects, each having over 1100 test frames (also see supplementary videos).…”

Unsupervised Feature Learning for RGB-D Based Object Recognition

“…The first dataset, 2D3D, consists of 156 object instances organized into 14 categories [10]. The authors of this dataset also use a large set of 2D and 3D manually designed shape and color features.…”

“…Two major innovations are introduced in this work: (1) Unsupervised feature learning on both color and depth channels; (2) spatial pyramid pooling over sparse codes from both layers of the HMP hierarchy. Extensive evaluations on several publicly available benchmark datasets [20,10,39] allowed us to gain various experimental insights: unsupervised feature learning from raw data can yield recognition accuracy that is superior to state-of-the-art object recognition algorithms, even to ones specifically designed and tuned for textured objects; the innovations introduced in this work significantly boost the performance of HMP applied to RGB-D data; and our approach can take full advantage of the additional information contained in color and depth channels.…”

A benchmark of simulated range images for partial shape retrieval

“…Table 1 shows some characteristics of previous benchmarks compared to our proposed dataset. It is also worth noting that previous efforts have been made to build benchmarks with scanned objects in the context of object recognition [5,11]. However, in those cases, the 3D object of the scanned object is often not available and therefore it is not possible to use algorithms based on the geometry of the target model.…”