🔍 Introduction

This is the official implementation for Object-IR (PR2026).

Tianli Liao¹, Ran Wang¹, Siqing Zhang¹, Lei Li¹, Guangen Liu¹, Chenyang Zhao¹, Heling Cao¹, Peng Li²

¹College of Information Science and Engineering, Henan University of Technology

²Institute for Complexity Science, Henan University of Technology

Pipeline

Given any aspect ratio, we construct a rigid mesh for the output resolution and estimate the grid's motion via a CNN-based regression network.

📝 Changelog

• 2026.05.13: Some codes are modified for better re-implementation.
• 2025.12.05: The training code, dataset and pretrained model are online.
• 2025.11.03: The paper of the arXiv version is online.

🐍 Code

Requirement

• python 3.8.5
• numpy 1.24.4
• pytorch 2.4.1
• tensorboard 2.13.0

We implement Object-IR with one GPU of RTX3090. Refer to environment.yml for more details.

Training

Step 1: Download the dataset

First download the COCO dataset (or in Baidu Cloud, extraction code: 1205). Unzip and put the dataset in the "Data/" directory.

Step 2: Start training

Run the following command from the project root to start training:

python Codes/train.py --train_path Data/train --batch_size 16 --max_epoch 50 --gpu 0

The trained model will be saved in the "model/" directory.

Pre-trained model

The pre-trained model are available at Google Drive or Baidu Cloud (extraction code: 1205). Please download them and put them in the 'model' folder.

Testing

Run the following commands from the project root. Relative paths such as Data/test and Codes/yolo11n.pt are resolved inside this repository.

Expected test data layout:

Data/test/
  input/      # test images
  output/     # generated by test_output.py, used by test.py

Step 1: Generate the warped images

Generate retargeted images from Data/test/input. By default, the warped images are saved into Data/test/output, so they can be consumed directly by test.py.

python Codes/test_output.py --test_path Data/test --output_path Data/test/output --grid_path grid --ratio_h 1 --ratio_w 0.5 --batch_size 1 --gpu 0

Useful arguments for test_output.py:

Argument	Default	Description
--test_path	Data/test	Test dataset root. It must contain an input/ folder.
--output_path	Data/test/output	Folder for generated warped images.
--grid_path	grid	Folder for mesh visualization images.
--ratio_h	1	Output height ratio.
--ratio_w	0.5	Output width ratio.
--batch_size	1	Batch size for inference.
--gpu	0	GPU id used by CUDA.

Step 2: Calculate the distortion error

Calculate the distortion error using Data/test/input and Data/test/output. The output/ folder is not a ground-truth label folder; it should contain the warped images generated in Step 1, with the same order and count as input/.

python Codes/test.py --test_path Data/test --ratio 0.5 --batch_size 1 --gpu 0 --yolo_path Codes/yolo11n.pt

Useful arguments for test.py:

Argument	Default	Description
--test_path	Data/test	Test dataset root. It must contain both input/ and output/.
--ratio	0.5	Output width ratio used when resizing outputs for the metric. Use the same value as --ratio_w when --ratio_h is 1.
--batch_size	1	Batch size for metric calculation.
--gpu	0	GPU id used by CUDA.
--yolo_path	Codes/yolo11n.pt	YOLO weight file used by the distortion metric.

📚 Citation

If you have any questions, please don't hesitate to contact me.

Tianli Liao -- [email protected]

@article{liao2026object-ir,
title = {Object-IR: Leveraging object consistency and mesh deformation for self-supervised image retargeting},
author = {Tianli Liao and Ran Wang and Siqing Zhang and Lei Li and Guangen Liu and Chenyang Zhao and Heling Cao and Peng Li},
journal = {Pattern Recognition},
volume = {172},
pages = {112651},
year = {2026},
}