Parsa Esmaeilkhani, Longin Jan Latecki · Temple University
Note: The code is to be completed.
git clone https://github.com/ParsaKhani/Angular_object_grounding.git
cd Angular_object_grounding
pip install -r requirements.txtTested with torch 2.5.1, transformers 5.10, peft 0.18, deepspeed 0.18.
Training data is a JSON list in LLaVA format. Image paths are relative to --image_folder:
[
{
"image": "train2014/COCO_train2014_000000581857.jpg",
"conversations": [
{"from": "human", "value": "<image> Please provide the bounding box coordinate of the region this sentence describes: the lady with the blue shirt."},
{"from": "gpt", "value": "[103.93, 299.99, 238.15, 477.41]"}
]
}
]Only the assistant turns are supervised.
bash scripts/train_qwen2_5_vl.shEdit the variables at the top of the script to set the data paths, the number of GPUs and the hyperparameters.
LoRA is applied to the language model, the vision-to-language projector is fully trained, and the vision encoder is frozen.
The LoRA adapter (with the projector weights) is saved to OUTPUT_DIR.
@article{esmaeilkhani2026angular,
title = {Angular Supervision for Visual Object Grounding in Autoregressive Vision-Language Models},
author = {Esmaeilkhani, Parsa and Latecki, Longin Jan},
journal = {arXiv preprint arXiv:2610.32220},
year = {2026}
}Apache-2.0. See LICENSE.