ConsNet

This repository maintains the official implementation of the paper ConsNet: Learning Consistency Graph for Zero‐Shot Human‐Object Interaction Detection by Ye Liu, Junsong Yuan and Chang Wen Chen, which has been accepted by ACM Multimedia 2020.

Installation

The ConsNet package could be installed directly from PyPI or manually from source for different uses. Please refer to the following environmental settings that we use.

• CUDA 10.2 Update 2
• CUDNN 8.0.5.39
• Python 3.9.2
• PyTorch 1.8.1
• MMDetection 2.11.0
• AllenNLP 2.2.0
• NNCore 0.2.4

Install from PyPI

You may install ConsNet from PyPI and import it in your own project as a Python package. This library implements several useful functionalities including Pair IoU, Pair NMS and unified APIs for HICO-DET dataset.

Simply run the following command to install the latest version of ConsNet.

pip install consnet

For more details about consnet.api, please refer to our documentation.

Install from source

By installing ConsNet from source, you may access the full capabilities of this project, including pooling object features, constructing the consistency graph and benchmarking the ConsNet model.

1. Clone the repository from GitHub.

git clone https://github.com/yeliudev/ConsNet.git
cd ConsNet

2. Install full dependencies and the package.

pip install -e .[full]

Getting Started

We pre-extract the visual features of all the humans and objects in the dataset and save them for training as well as testing. These features are also used to construct the consistency graph. Please refer to our paper for more details about feature extraction and data sampling.

Build dataset and construct the consistency graph

1. Download the checkpoints of object detector and ELMo.

ROOT='https://s3-us-west-2.amazonaws.com/allennlp/models/elmo'
ELMO='2x4096_512_2048cnn_2xhighway_5.5B'

# Download object detector checkpoints
wget https://dl.catcatdev.com/consnet/faster_rcnn_r50_fpn_3x_coco-26df6f6b.pth
wget https://dl.catcatdev.com/consnet/faster_rcnn_r50_fpn_20e_hico_det-77b91312.pth

# Download ELMo options and weights
wget ${ROOT}/${ELMO}/elmo_${ELMO}_options.json
wget ${ROOT}/${ELMO}/elmo_${ELMO}_weights.hdf5

2. Download HICO-DET dataset and prepare the files in the following structure.

ConsNet
├── configs
├── consnet
├── tools
├── checkpoints
│   ├── faster_rcnn_r50_fpn_3x_coco-26df6f6b.pth
│   ├── faster_rcnn_r50_fpn_20e_hico_det-77b91312.pth
│   ├── elmo_2x4096_512_2048cnn_2xhighway_5.5B_options.json
│   └── elmo_2x4096_512_2048cnn_2xhighway_5.5B_weights.hdf5
├── data
│   └── hico_20160224_det
│       ├── anno_bbox.mat
│       └── images
│           ├── train2015
│           └── test2015
├── README.md
├── setup.py
└── ···

3. Convert the annotations to COCO format. The results will be saved in data/hico_det/annotations.

python tools/convert_anno.py

4. Build dataset and construct the consistency graph. The results will be saved in data/hico_det.

python tools/build_dataset.py --checkpoint <path-to-checkpoint>

Train a model

Run the following command to train a model using specified configs.

python tools/launch.py --config <path-to-config>

Test a model and evaluate results

Run the following command to test a model and evaluate results.

python tools/launch.py --config <path-to-config> --checkpoint <path-to-checkpoint> --eval

Model Zoo

We provide multiple HICO-DET pre-trained models here. All the models are trained using a single NVIDIA Tesla V100-SXM2 GPU and are evaluated under the default metric of HICO-DET dataset.

Detector	Model	Type	Performance (mAP)					Download
			Full	Rare	Non-Rare	Seen	Unseen
COCO	ConsNet	UC	19.78	14.43	21.37	20.69	16.13	model | metrics
	ConsNet	UO	20.71	16.81	21.87	20.99	19.27	model | metrics
	ConsNet	UA	19.04	14.54	20.38	20.02	14.12	model | metrics
	ConsNet	GT	53.04	38.79	57.3	—	—	model | metrics
	ConsNet	—	22.15	17.55	23.52	—	—
HICO-DET	ConsNet-F	—	25.94	19.35	27.91	—	—	model | metrics

Note that: Type UC, UO, UA and GT represent unseen action-object combination, unseen object, unseen action and ground truth scenarios respectively.

Customization

Thanks to the modulized implementation based on NNCore, this project is highly customizable with a number of replaceable modules. You may play with the hyperparameters in configs or construct your own HOI detection pipeline by replacing the dataset, detector, embedder, etc. Please check the documentation of NNCore for more details about customizing the engine and modules.

Citation

If you find this project useful for your research, please kindly cite our paper.

@inproceedings{liu2020consnet,
  title={ConsNet: Learning Consistency Graph for Zero-Shot Human-Object Interaction Detection},
  author={Liu, Ye and Yuan, Junsong and Chen, Chang Wen},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  pages={4235--4243},
  year={2020}
}