An RGB-D Image Dataset for Lychee Detection and Maturity Classification for Robotic Harvesting
Description
The dataset contains color (RGB) images collected under different weather conditions and at different time periods, with a resolution of 1280×1024. The images cover various lychee varieties, such as Nuomi, Feizi Xiao, Heiye, and Huaizhi. The dataset includes three distinct ripening stages: immature, semi-ripe, and mature, comprising a total of 11,414 images. These include 878 original RGB images, 8,780 enhanced RGB images, and 1,756 depth images. The images are organized into five folders based on collection date, such as Image_YoloLabel_250605_Outdoor_sunny. Each folder includes descriptions of the weather and indoor/outdoor scenes. Each image and its labeled file is numbered, and includes the data augmentation method, number of categories, and similarity score (e.g., ID-0001_cc_unripe-0_semi-ripe-0_ripe-2_sim-0.0016.jpg). These images are labeled with 9,658 pairs of tags for lychee detection and ripeness classification during robotic harvesting. To improve the consistency of annotation, three researchers independently annotated the data, and then a fourth reviewer summarized and verified their results.
Files
Steps to reproduce
1.we constructed a dataset to facilitate lychee detection and maturity classification, which was collected using a self-developed multi-modal sensor module. 2.Videos were captured from one side of lychee trees at a distance of roughly 20–60 cm from the canopies, with a walking speed of approximately 1 m/s. The low-speed movement helped reduce motion-induced image blur. Data collections were conducted during the three-week peak of lychee ripening (specifically, on June 5, 10, 11, 12, and 19, 2025). Multiple common lychee varieties, such as Nuomici, Feizixiao, Heiye, and Huaizhi, were captured under various weather conditions (e.g., light rain on June 11, and sunny on other days) and at different times of day (morning, noon, and evening). The ROS plugin (bag_to_images) was used to convert the recorded .bag files into image sequences. One image was extracted for every 10 frames in the video. To improve dataset diversity and reduce redundancy, after the first image was manually selected in a sequence, the Structural Similarity Inde was employed to select the most dissimilar frames within each 10-frame segment from the previous selected image.
Institutions
- Shenzhen University