Erdut dataset

Erdut dataset is a challenging fruit detection dataset containing RGB-D images of four fruit sorts acquired in a highly unstructured environment and in uncontrolled lighting conditions and includes a semi automatically annotated ground truth data. With the aid of a hand held PrimeSense Carmine 1.09 short range sensor operating in 100mm-mode, several sequences of RGB-D images of resolution 320 × 240 were captured in natural outdoor conditions in an orchard. Each sequence contained images of one of the following four fruits: tomato, nectarine, pear or plum. For experimental purposes, images from these sequences were selected using two criteria. The first criterion was that there was no motion blur or misalignment of depth and RGB image. Such images were obtained when the camera was steady or moving very slowly. Since the image segments needed to be manually labelled for classifier training and testing, the second criterion required that the fruit in the image was not deep in a shadow so that it could be visually unambiguously categorized by a human evaluator. Each of the selected images was then segmented into approximately convex surfaces using the procedure described in [1] and each segment of image was manually labelled by a human evaluator. The segmentation procedure was configured to consider only the points in the depth image within a distance of 0.6 m from the camera. The images and corresponding approximately convex segments were then divided into two datasets: a training or reference dataset, which was used for classifier training, and a test dataset used to assess the discriminatory quality of each of the compared descriptors. Examples of RGB images contained in the dataset are shown below.

Download Dataset