Hybrid-SOV

Hybrid-SOV

Bridging Detection Architectures With Foundation Models: A Unified Framework for Human-Object Interaction Detection.

[Paper]

Requirements

conda create -n hybrid-sov python=3.11 -y
conda activate hybrid-sov
pip install uv

# Install PyTorch/torchvision for your CUDA version first.
uv pip install torch torchvision

uv pip install numpy scipy pillow tqdm pyyaml pycocotools tabulate addict yapf loguru
uv pip install timm transformers fairscale omegaconf wandb
uv pip install git+https://github.com/openai/CLIP.git
uv pip install -U huggingface_hub

Dataset Preparation

HICO-DET

Please follow the HICO-DET dataset preparation of GGNet. See also the README of QAHOI.

After preparation, the data/hico_det folder should look like:

data
+-- hico_det
|   +-- images
|   |   +-- test2015
|   |   +-- train2015
|   +-- annotations
|       +-- anno_list.json
|       +-- corre_hico.npy
|       +-- file_name_to_obj_cat.json
|       +-- hoi_id_to_num.json
|       +-- hoi_list_new.json
|       +-- test_hico.json
|       +-- trainval_hico.json

V-COCO

Please follow the installation of V-COCO.

For evaluation, please put vcoco_test.ids and vcoco_test.json into the data/v-coco/data folder.

After preparation, the data/v-coco folder should look like:

data
+-- v-coco
|   +-- prior.pickle
|   +-- images
|   |   +-- train2014
|   |   +-- val2014
|   +-- data
|   |   +-- instances_vcoco_all_2014.json
|   |   +-- vcoco_test.ids
|   |   +-- vcoco_test.json
|   +-- annotations
|       +-- corre_vcoco.npy
|       +-- test_vcoco.json
|       +-- trainval_vcoco.json

Model Weights

The checkpoint files are under the params/ folder in the Hugging Face repository.

Download the released weights into the local params folder:

huggingface-cli download thxplz/Hybrid-SOV \
  --include "params/*" \
  --local-dir .

Expected files include:

params
+-- hico_det_hybrid-sov-r50.pth
+-- hico_det_hybrid-sov-vla-r50.pth
+-- rtdetr_r50vd_6x_coco_from_paddle_converted_hico.pth
+-- rtdetr_r50vd_6x_coco_from_paddle_converted_vcoco.pth
+-- vcoco-hybrid-sov-vla-r50.pth

Evaluation

HICO-DET

Evaluate the released checkpoints by running:

# Hybrid-SOV-R50 (HICO-DET)
sh run/hybrid-sov-r50_eval.sh

# Hybrid-SOV-VLA-R50 (HICO-DET)
sh run/hybrid-sov-vla-r50_eval.sh

V-COCO

Evaluate the released checkpoint by running:

# Hybrid-SOV-VLA-R50 (V-COCO)
sh run/vcoco-hybrid-sov-vla-r50_eval.sh

Training

HICO-DET

Download the RT-DETR R50 HICO pre-trained weight from thxplz/Hybrid-SOV into params:

params/rtdetr_r50vd_6x_coco_from_paddle_converted_hico.pth

Train Hybrid-SOV-R50:

sh run/hybrid-sov-r50.sh

Train Hybrid-SOV-VLA-R50:

sh run/hybrid-sov-vla-r50.sh

Train Hybrid-SOV-VLA with DINOv3/ConvNeXt-L:

sh run/hybrid-sov-vla-dinov3-cnx-l.sh

V-COCO

Download the RT-DETR R50 V-COCO pre-trained weight from thxplz/Hybrid-SOV into params:

params/rtdetr_r50vd_6x_coco_from_paddle_converted_vcoco.pth

Train Hybrid-SOV-VLA-R50 on V-COCO:

sh run/vcoco-hybrid-sov-vla-r50_hoi.sh

References

@ARTICLE{chen2026hybridsov,
  author={Chen, Junwen and Yanai, Keiji},
  journal={IEEE Access}, 
  title={Bridging Detection Architectures With Foundation Models: A Unified Framework for Human-Object Interaction Detection}, 
  year={2026},
  volume={14},
  pages={23299-23310},
  url={https://ieeexplore.ieee.org/document/11367687}
}