1611 strawberries from different places and varieties are used to collect images, 3D scans as well as physical properties such as shape, width, height, and weight.

The database described in this paper was made using 1611 strawberries divided into 20 groups, representing 15 varieties and harvested at 3 locations: Oxnard, Santa Maria and Watsonville (California, USA). A description of each group is given in table \ref{tab:straw}. The strawberries were harvested between June and September 2014. They were shipped overnight and delivered at the University of California, Davis the next day, where they were stored in a temperature-controlled chamber set at two degrees Celsius. They were processed within one up to four days after being picked.

The first step consisted in collecting 22 images for each berry with a NIKON DS 2000 camera. To do so, each strawberry was placed on a robotic arm and rotated between each picture. For images 1 to 11, the robot rotates around vector x by 18 degrees. For images 12 to 22, the rotation is around vector z by 18 degrees. In order to obtain a uniform set of images, the lighting environment was controlled and the white balance was held constant over time. 

The second step of the data collection consisted in identifying the shape of each strawberry, measure its dimension, and weigh it with and without the calyx. First, the shape classification was manually done. Next, the maximal height and width were manually measured with a caliper. Finally, each fruit was weighed with and without the calyx thanks to a Scout Pro SP602 scale. The calyx was manually removed while trying to minimize the amount of removed flesh.

The final step consisted of scanning the flesh of the strawberries. The scanning was performed using a Solutionix Rexcan DS2 scanner. The fruits were spray-coated with a white titanium solution to minimize light reflection. Each berry was then pinned on a mobile platform within the scanner. Finally, 10 different views were used to create the 3D model. 

For each fruit, the scanning process provides a point cloud, which was then manually processed using the Meshlab software. First, the close vertices were merged to reduce the point cloud size. Next, the points corresponding to the surface of the pin were removed, and finally, a Poisson surface reconstruction filter was applied to the point cloud using a tree depth of 4, 6 and 8.