YOLO-based models trained via transfer learning with top-view person image annotations.

This dataset includes YOLOv5, YOLOv8, and YOLOv11 models fine-tuned via transfer learning on top-view (zenithal) person images.

YOLO models customized for detecting and classifying the person class in top-down (zenithal) views. The original YOLO architectures were adapted by disabling the detection of frontal or lateral human views and removing all other object classes recognized by the base models. The training dataset consists of 3,137 manually annotated images collected from three university libraries, capturing a rich variety of top-view human patterns, including individuals alone, overlapping people, partial occlusions, and diverse clothing styles.

Transfer learning was applied to YOLO versions v5, v8, and v11, using both their nano variants (optimized for low computational cost) and xlarge variants (providing higher accuracy at the expense of greater computational demand).

It also provides a Python script (.py) to run each model on sample images for person detection.
These sample images were not used during the training or validation phases.

File List (if file is compressed, list its content): 
/images → A test set of .jpg images depicting people from a top-view camera perspective.

models → a set of YOLO models in .pt format, including versions 5, 8, and 11  ( _n nano,  _x xtralarge)

yolov5_n.pt

yolov8_n.pt

yolov11_n.pt

yolov5_x.pt

yolov8_x.pt

yolov11_x.pt

Yolo_test.py → A Python script demonstrating how to use a YOLO model to detect people in the sample images. You can edit the script to select the model and specify the paths to the image folder and the model files. 

File format (provide a list of all file formats present in this dataset):
.jpg → image files

.pt → YOLO model files

.py → Python script file