youngfly11 / promting_hoi

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Transferable Human-object Interaction Detector (THID)

Overview

THID is an end-to-end transformer-based human-object interaction (HOI) detector. [Paper]

THID

  • Motivation: It is difficult to construct a data collection including all possible combinations of human actions and interacting objects due to the combinatorial nature of human-object interactions (HOI). In this work, we aim to develop a transferable HOI detector for the wide range of unseen interactions.
  • Components: (1) We treat independent HOI labels as the natural language supervision of interactions and embed them into a joint visual-and-text space to capture their correlations. (2) Our visual encoder is instantiated as a Vision Transformer with new learnable HOI tokens and a sequence parser to generate HOI predictions with bounding boxes. (3) It distills and leverages the transferable knowledge from the pretrained CLIP model to perform the zero-shot interaction 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 --use_env main.py \
    --batch_size 8 \
    --output_dir [path to save checkpoint] \
    --epochs 100 \
    --lr 1e-4 --min-lr 1e-7 \
    --hoi_token_length 50 \
    --enable_dec \
    --dataset_file hico

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

python -m torch.distributed.launch --nproc_per_node=2 --use_env main.py \
    --batch_size 8 \
    --output_dir [path to save checkpoint] \
    --epochs 100 \
    --lr 1e-4 --min-lr 1e-7 \
    --hoi_token_length 50 \
    --enable_dec \
    --dataset_file swig

Inference

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

python main.py --eval \
    --batch_size 1 \
    --output_dir [path to save results] \
    --hoi_token_length 50 \
    --enable_dec \
    --pretrained [path to the pretrained model] \
    --eval_size 256 [or 224 448 ...] \
    --test_score_thresh 1e-4 \
    --dataset_file hico

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

python main.py --eval \
    --batch_size 8 \
    --output_dir [path to save results] \
    --hoi_token_length 10 \
    --enable_dec \
    --pretrained [path to the pretrained model] \
    --eval_size 256 [or 224 448 ...] \
    --test_score_thresh 1e-4 \
    --dataset_file swig

Models

Model dataset HOI Tokens AP seen AP unseen Log Checkpoint
THID-HICO HICO-DET 50 25.30 17.57 Log params
THID-HICO HICO-DET 10 23.72 16.45 Log params
Model dataset HOI Tokens AP non-rare AP rare AP unseen Log Checkpoint
THID-SWIG SWIG-HOI 20 19.49 14.13 10.49 Log params
THID-SWIG SWIG-HOI 10 18.30 13.99 11.14 Log params

Citing

Please consider citing our paper if it helps your research.

@inproceedings{wang_cvpr2022,
author = {Wang, Suchen and Duan, Yueqi and Ding, Henghui and Tan, Yap-Peng and Yap, Kim-Hui and Yuan, Junsong},
title = {Learning Transferable Human-Object Interaction Detectors with Natural Language Supervision},
booktitle = {CVPR},
year      = {2022},
}

About

License:MIT License


Languages

Language:Python 100.0%