Data, models and code for "Background Suppression with Plant-focused Image Transformation for Robust Plant Detection in Field Images"

This record contains the data, the trained models and the code of the paper "Background Suppression with Plant-focused Image Transformation for Robust Plant Detection in Field Images". Plant-focused Image Transformation (PIT) is a preprocessing method that keeps the target plant in a field image and sets the background to black. Six object detectors (Faster R-CNN, SSD, EfficientDet-D1, YOLOv8s, YOLO26s and RT-DETR-l) were trained on raw images and on PIT images and tested on the original test set, on a background-swapped test set and on plant-free soil images. The record contains:- 2444 field images (640 x 640) of five species (Rhaponticum repens, Panicum miliaceum, Salsola kali, Sinapis arvensis, Zea mays) with boxes, PIT images, PIT masks and 100 manually corrected masks- the 12 trained detectors and their predictions on the three test sets- the training settings and training curves of all 12 detectors- the code of PIT, of the evaluation and of the training; one script recomputes Tables 1, 4, 5 and S1-S3 of the paper Extract data.zip, models.zip and results.zip in the same folder as the scripts. See README.md for the folder structure and the commands. Licenses: code MIT; data and TensorFlow models CC BY 4.0; Ultralytics models (YOLOv8s, YOLO26s, RT-DETR-l; fine-tuned from Ultralytics weights) AGPL-3.0. See LICENSE.