Dataset and Reproducibility Resources for Object Detection in Pepper Harvesting

This record provides the dataset and reproducibility resources associated with a comparative evaluation of object detection methods for multiclass visual perception in pepper-harvesting environments. The dataset comprises 1004 RGB images and 4445 bounding-box annotations for five object categories: human, pepper, pepper group, box, and pushcart. The collection combines 704 field-acquired photographs obtained in Chile with 300 complementary images selected from two publicly available pepper-related datasets: 198 images from the Murat dataset and 102 images from the Khai dataset. All retained public-source images were manually re-annotated using the unified five-class annotation scheme adopted in this study. The fixed partition contains 803 training, 100 validation, and 101 test images. The deposited resources preserve the dataset partition used in the associated experiments and include supporting integrity-audit and reproducibility materials. The final dataset audit identified no exact cross-split duplicate groups and no potential cross-split near-duplicate pairs under the reported dHash criterion.