[CVPR2025] Locality-Aware Zero-Shot Human-Object Interaction Detection

This repo is the official code for our CVPR2025 paper "Locality-Aware Zero-Shot Human-Object Interaction Detection".

Installation

Follow the instructions to install dependencies.

git clone git@github.com:OreoChocolate/LAIN.git

conda create -n lain python=3.8.18
conda activate lain   
conda install pytorch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1  pytorch-cuda=11.8 -c pytorch -c nvidia
pip  matplotlib, tqdm, scipy, regex, ftfy, wandb, gdown

cd pocket
pip install -e .

Data preparation

Dataset download

We provide a shell script for downloading the dataset. You can download all the required files by running the following command:

bash scripts/download.sh

Model Zoo

Zero-shot setting	Backbone	Unseen	Seen	mAP
RF_UC	ResNet50+ViT-B/16	32.34	35.17	34.60
NF_UC	ResNet50+ViT-B/16	36.67	32.63	33.44
UO	ResNet50+ViT-B/16	37.65	33.61	34.28
UV	ResNet50+ViT-B/16	29.23	33.95	33.29

You can download the pretrained models from this link.

The downloaded files should be placed as follows:

LAIN
├── hicodet
│   └── hico_20160224_det
└── checkpoints 
    ├── pretrained_clip
    └── pretrained_detr

Training & Evaluation

We provide four training and evaluation scripts for different zero-shot HOI detection settings (i.e., RF-UC, NF-UC, UO and UV), all located in the scripts folder.

#training
bash scripts/training/[NF-UC,RF-UC,UO,UV].sh

#evaluation
bash scripts/eval/[NF-UC,RF-UC,UO,UV].sh

Acknowledgement

Our implementation is built upon ADA-CM and UPT. We are grateful to the authors for their excellent work and for making their code publicly available.