Request for YOLOv12 Architecture & Training Data Details for Research Reproduction

Discussion
, ,
1

Hello Ultralytics Team,

I am conducting a research study where I aim to:

  1. Reproduce the YOLOv12 training process,

  2. Analyze architectural variations, and

  3. Evaluate how modifications affect performance and generalization.

To do this accurately, I wanted to ask:

1. Is detailed documentation of the YOLOv12 architecture available?

Including layer configurations, architectural modifications compared to YOLOv8/YOLOv10, training objectives, or any design choices made during development.

2. Is the original training dataset (or dataset composition details) publicly available?

If not, any information on dataset type, preprocessing, augmentation pipeline, or data distribution would be very helpful for reproducing results.

This project is strictly for research and academic use, and I’m happy to share any findings or insights that may contribute back to the community. Proper attribution will be given in all publications.

2
  1. That’s probably found in the YOLO12 paper: [2502.12524] YOLOv12: Attention-Centric Real-Time Object Detectors
  2. Models are trained on MS COCO dataset which is public. The command to train the model is shown in the author’s repo: yolov12/README.md at 3bca22b336e96cfdabfec4c062b84eef210e9563 · sunsmarterjie/yolov12 · GitHub