CMD-SE

This repository contains the official PyTorch implementation for the paper:

Ting Lei, Shaofeng Yin, Yang Liu; Exploring the Potential of Large Foundation Models for Open-Vocabulary HOI Detection; In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024

Overview

Open-vocabulary human-object interaction (HOI) detection, which is concerned with the problem of detecting novel HOIs guided by natural language, is crucial for understanding human-centric scenes. However, prior zero-shot HOI detectors often employ the same levels of feature maps to model HOIs with varying distances, leading to suboptimal performance in scenes containing human-object pairs with a wide range of distances. In addition, these detectors primarily rely on category names and overlook the rich contextual information that language can provide, which is essential for capturing open vocabulary concepts that are typically rare and not well-represented by category names alone. In this paper, we introduce a novel end-to-end open vocabulary HOI detection framework with conditional multi-level decoding and fine-grained semantic enhancement~(CMD-SE), harnessing the potential of Visual-Language Models (VLMs). Specifically, we propose to model human-object pairs with different distances with different levels of feature maps by incorporating a soft constraint during the bipartite matching process. Furthermore, by leveraging large language models (LLMs) such as GPT models, we exploit their extensive world knowledge to generate descriptions of human body part states for various interactions. Then we integrate the generalizable and fine-grained semantics of human body parts to improve interaction recognition. Experimental results on two datasets, SWIG-HOI and HICO-DET, demonstrate that our proposed method achieves state-of-the-art results in open vocabulary HOI detection.

Preparation

Installation

Our code is built upon CLIP. This repo requires to install PyTorch and torchvision, as well as small additional dependencies.

conda install pytorch torchvision cudatoolkit=11.3 -c pytorch
pip install ftfy regex tqdm numpy Pillow matplotlib

Dataset

The experiments are mainly conducted on HICO-DET and SWIG-HOI dataset. We follow this repo to prepare the HICO-DET dataset. And we follow this repo to prepare the SWIG-HOI dataset.

HICO-DET

HICO-DET dataset can be downloaded here. After finishing downloading, unpack the tarball (hico_20160224_det.tar.gz) to the data directory. We use the annotation files provided by the PPDM authors. We re-organize the annotation files with additional meta info, e.g., image width and height. The annotation files can be downloaded from here. The downloaded files have to be placed as follows. Otherwise, please replace the default path to your custom locations in datasets/hico.py.

 |─ data
 │   └─ hico_20160224_det
 |       |- images
 |       |   |─ test2015
 |       |   |─ train2015
 |       |─ annotations
 |       |   |─ trainval_hico_ann.json
 |       |   |─ test_hico_ann.json
 :       :

SWIG-DET

SWIG-DET dataset can be downloaded here. After finishing downloading, unpack the images_512.zip to the data directory. The annotation files can be downloaded from here. The downloaded files to be placed as follows. Otherwise, please replace the default path to your custom locations in datasets/swig.py.

 |─ data
 │   └─ swig_hoi
 |       |- images_512
 |       |─ annotations
 |       |   |─ swig_train_1000.json
 |       |   |- swig_val_1000.json
 |       |   |─ swig_trainval_1000.json
 |       |   |- swig_test_1000.json
 :       :

Training

Run this command to train the model in HICO-DET dataset

python -m torch.distributed.launch --nproc_per_node=2 --master_port 3990 --use_env main.py \
    --batch_size 64 \
    --output_dir [path to save checkpoint] \
    --epochs 80 \
    --lr 1e-4 --min-lr 1e-7 \
    --hoi_token_length 25 \
    --enable_dec \
    --dataset_file hico --multi_scale true --f_idxs 5 8 11 --set_cost_hoi_type 5 --use_aux_text true \
    --enable_focal_loss --description_file_path hico_hoi_descriptions.json

Run this command to train the model in SWIG-HOI dataset

python -m torch.distributed.launch --nproc_per_node=2 --master_port 4990 --use_env main.py \
    --batch_size 64 \
    --output_dir [path to save checkpoint] \
    --epochs 70 \
    --lr 1e-4 --min-lr 1e-7 \
    --hoi_token_length 10 \
    --enable_dec \
    --dataset_file swig --multi_scale true --f_idxs 5 8 11 --set_cost_hoi_type 5 --use_aux_text true \
    --enable_focal_loss --description_file_path swig_hoi_descriptions_6bodyparts.json

Inference

Run this command to evaluate the model on HICO-DET dataset

python -m torch.distributed.launch --nproc_per_node=1 --master_port 3990 --use_env main.py \
    --batch_size 64 \
    --output_dir [path to save checkpoint] \
    --epochs 80 \
    --lr 1e-4 --min-lr 1e-7 \
    --hoi_token_length 25 \
    --enable_dec \
    --dataset_file hico --multi_scale true --f_idxs 5 8 11 --set_cost_hoi_type 5 --use_aux_text true \
    --enable_focal_loss --description_file_path hico_hoi_descriptions.json \
    --eval --pretrained [path to ckpt]

Run this command to evaluate the model on SWIG-HOI dataset

python -m torch.distributed.launch --nproc_per_node=1 --master_port 4990 --use_env main.py \
    --batch_size 64 \
    --output_dir [path to save results] \
    --epochs 70 \
    --lr 1e-4 --min-lr 1e-7 \
    --hoi_token_length 10 \
    --enable_dec \
    --dataset_file swig --multi_scale true --f_idxs 5 8 11 --set_cost_hoi_type 5 --use_aux_text true \
    --enable_focal_loss --description_file_path swig_hoi_descriptions_6bodyparts.json \
    --eval --pretrained [path to ckpt]

Models

dataset	HOI Tokens	Unseen	Seen	Full	Checkpoint
HICO-DET	25	16.70	23.95	22.35	params

dataset	HOI Tokens	Non-rare	Rare	Unseen	Full	Checkpoint
SWIG-HOI	10	21.46	14.64	10.70	15.26	params

Citing

Please consider citing our paper if it helps your research.

@inproceedings{CMD-SE_2024_CVPR,
 title={Exploring the Potential of Large Foundation Models for Open-Vocabulary HOI Detection},
 author={Ting Lei, Shaofeng Yin, and Yang Liu},
 year={2024},
 booktitle={CVPR},
 organization={IEEE}
}

Acknowledgement

We thank THID for open-sourcing their code. We would also like to thank the anonymous reviewers for their constructive feedback.