Research Article
Mitigation of Effects of Occlusion on Object Recognition with Deep Neural Networks through Low-Level Image Completion
Figure 1
Sample images from NORB, ImageNet, and SORBO datasets. NORB (a) is a parametric dataset designed for experiments in invariant object classification. It includes five categories of objects, each with ten specific objects. NORB contains a stereo pair at nine camera elevations, eighteen camera azimuths, and five light levels for each instance [12]. ImageNet (b) is a nonparametric dataset containing a large number of labeled examples scraped from the Internet. In comparison to NORB, ImageNet has much more data and many more categories but typically only a single image for each object instance. The image parameters are unknown except for the category [7]. SORBO (c) is a new dataset that extends the NORB data. It preserves the rich parametric metadata from NORB but adds various levels of bar, blob, and random stereo occlusions.
(a) |
(b) |
(c) |