RGB-D Time-Series Dataset of White Button Mushroom Growth for Instance Segmentation

This dataset was collected to support research on computer vision–based perception systems for autonomous mushroom harvesting. A key challenge in robotic harvesting of white button mushrooms (Agaricus bisporus) is the accurate detection and segmentation of individual mushroom caps in densely populated cultivation beds. Precise boundary delineation is required for estimating cap size, determining grasp points, and enabling collision-free manipulation in automated harvesting systems.

RGB and depth images were collected from top-view perspectives over indoor mushroom production beds. The dataset was acquired using a rail-mounted RGB–D imaging system equipped with a ZED Mini stereo camera positioned approximately 749 mm above the mushroom beds. Because the beds exceeded the camera’s field of view, images were captured at multiple overlapping positions along the rail and later combined to form stitched mosaics representing the entire bed area.

Images were collected every 10 minutes. At each time point, 17 partially overlapping RGB–D frames were captured and stitched to generate a single mosaic image of the cultivation beds. In total, the dataset contains 129 stitched RGB–depth mosaics representing time-series observations of mushroom growth and spatial distribution.

The dataset includes three primary components: RGB images, corresponding depth maps, and pixel-level annotations of individual mushroom caps. The annotations provide instance segmentation masks outlining the boundaries of each mushroom cap and are provided in COCO format. Depth data enables additional geometric analysis such as estimating mushroom elevation, cap curvature, and spatial relationships between neighboring mushrooms.

This dataset can be used to develop and evaluate perception algorithms for applications including agricultural robotics, automated mushroom harvesting, crop monitoring, and RGB–D scene understanding in dense agricultural environments. Researchers can use the dataset to train machine learning models, benchmark segmentation performance, analyze mushroom growth dynamics, and investigate perception strategies for robotic manipulation in controlled-environment agriculture.