Panoptic Scene Graph Generation

Panoptic Scene Graph Generation
Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, Ziwei Liu
S-Lab, Nanyang Technological University & SenseTime Research

Updates

Oct 31, 2022: We release the full dataset here that we used in the paper. The competition version is here.
Oct 9, 2022: The preliminary round of PSG challenge ends. We will release the entire dataset after the final round starts. Before that, if you want to access the PSG dataset (competition version), please email me.
Sep 4, 2022: We introduce the PSG Classification Task for NTU CE7454 Coursework, as described here.
Aug 21, 2022: We provide guidance on PSG challenge registration here.
Aug 12, 2022: Replicate demo and Cloud API is added, try it here!
Aug 10, 2022: We launched Hugging Face demo 🤗. Try it with your scene!
Aug 5, 2022: The PSG Challenge will be available on International Algorithm Case Competition ! All the data will be available there then! Stay tuned!
July 25, 2022: 💥 We are preparing a PSG competition with ECCV'22 SenseHuman Workshop and International Algorithm Case Competition, starting from Aug 6, with a prize pool of 🤑 US$150K 🤑. Join us on our Slack to stay updated!
July 25, 2022: PSG paper is available on arXiv.
July 3, 2022: PSG is accepted by ECCV'22.

What is PSG Task?

The Panoptic Scene Graph Generation (PSG) Task aims to interpret a complex scene image with a scene graph representation, with each node in the scene graph grounded by its pixel-accurate segmentation mask in the image.

To promote comprehensive scene understanding, we take into account all the content in the image, including "things" and "stuff", to generate the scene graph.


PSG Task: To generate a scene graph that is grounded by its panoptic segmentation

PSG addresses many SGG problems

We believe that the biggest problem of classic scene graph generation (SGG) comes from noisy datasets. Classic scene graph generation datasets adopt a bounding box-based object grounding, which inevitably causes a number of issues:

Coarse localization: bounding boxes cannot reach pixel-level accuracy,
Inability to ground comprehensively: bounding boxes cannot ground backgrounds,
Tendency to provide trivial information: current datasets usually capture frivolous objects like head to form trivial relations like person-has-head, due to too much freedom given during bounding box annotation.
Duplicate groundings: the same object could be grounded by multiple separate bounding boxes.

All of the problems above can be easily addressed by the PSG dataset, which grounds the objects using panoptic segmentation with an appropriate granularity of object categories (adopted from COCO).

In fact, the PSG dataset contains 49k overlapping images from COCO and Visual Genome. In a nutshell, we asked annotators to annotate relations based on COCO panoptic segmentations, i.e., relations are mask-to-mask.


Comparison between the classic VG-150 and PSG.

Clear Predicate Definition

We also find that a good definition of predicates is unfortunately ignored in the previous SGG datasets. To better formulate PSG task, we carefully define 56 predicates for PSG dataset. We try hard to avoid trivial or duplicated relations, and find that the designed 56 predicates are enough to cover the entire PSG dataset (or common everyday scenarios).

Type	Predicates
Positional Relations (6)	over, in front of, beside, on, in, attached to.
Common Object-Object Relations (5)	hanging from, on the back of, falling off, going down, painted on.
Common Actions (31)	walking on, running on, crossing, standing on, lying on, sitting on, leaning on, flying over, jumping over, jumping from, wearing, holding, carrying, looking at, guiding, kissing, eating, drinking, feeding, biting, catching, picking (grabbing), playing with, chasing, climbing, cleaning (washing, brushing), playing, touching, pushing, pulling, opening.
Human Actions (4)	cooking, talking to, throwing (tossing), slicing.
Actions in Traffic Scene (4)	driving, riding, parked on, driving on.
Actions in Sports Scene (3)	about to hit, kicking, swinging.
Interaction between Background (3)	entering, exiting, enclosing (surrounding, warping in)

Get Started

To setup the environment, we use conda to manage our dependencies.

Our developers use CUDA 10.1 to do experiments.

You can specify the appropriate cudatoolkit version to install on your machine in the environment.yml file, and then run the following to create the conda environment:

conda env create -f environment.yml

You shall manually install the following dependencies.

# Install mmcv
## CAUTION: The latest versions of mmcv 1.5.3, mmdet 2.25.0 are not well supported, due to bugs in mmdet.
pip install mmcv-full==1.4.3 -f https://download.openmmlab.com/mmcv/dist/cu101/torch1.7.0/index.html

# Install mmdet
pip install openmim
mim install mmdet=2.20.0

# Install coco panopticapi
pip install git+https://github.com/cocodataset/panopticapi.git

# For visualization
conda install -c conda-forge pycocotools
pip install detectron2==0.5 -f \
  https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.7/index.html

# If you're using wandb for logging
pip install wandb
wandb login

# If you develop and run openpsg directly, install it from source:
pip install -v -e .
# "-v" means verbose, or more output
# "-e" means installing a project in editable mode,
# thus any local modifications made to the code will take effect without reinstallation.

Datasets and pretrained models are provided. Please unzip the files if necessary.

Before October 2022, we only release part of the PSG data for competition, where part of the test set annotations are wiped out. Users should change the json filename in psg.py (Line 4-5) to a correct filename for training or submission.

For the PSG competition, we provide psg_train_val.json (45697 training data + 1000 validation data with GT). Participant should use psg_val_test.json (1000 validation data with GT + 1177 test data without GT) to submit. Example submit script is here. You can use grade.sh to simulate the competition's grading mechanism locally.

Our codebase accesses the datasets from ./data/ and pretrained models from ./work_dirs/checkpoints/ by default.

├── ...
├── configs
├── data
│   ├── coco
│   │   ├── panoptic_train2017
│   │   ├── panoptic_val2017
│   │   ├── train2017
│   │   └── val2017
│   └── psg
│       ├── psg_train_val.json
│       ├── psg_val_test.json
│       └── ...
├── openpsg
├── scripts
├── tools
├── work_dirs
│   ├── checkpoints
│   ├── psgtr_r50
│   └── ...
├── ...

We suggest our users to play with ./tools/Visualize_Dataset.ipynb to quickly get familiar with PSG dataset.

To train or test PSG models, please see https://github.com/Jingkang50/OpenPSG/tree/main/scripts for scripts of each method. Some example scripts are below.

Training

# Single GPU for two-stage methods, debug mode
PYTHONPATH='.':$PYTHONPATH \
python -m pdb -c continue tools/train.py \
  configs/psg/motif_panoptic_fpn_r50_fpn_1x_sgdet_psg.py

# Multiple GPUs for one-stage methods, running mode
PYTHONPATH='.':$PYTHONPATH \
python -m torch.distributed.launch \
--nproc_per_node=8 --master_port=29500 \
  tools/train.py \
  configs/psgformer/psgformer_r50_psg.py \
  --gpus 8 \
  --launcher pytorch

Testing

# sh scripts/imp/test_panoptic_fpn_r50_sgdet.sh
PYTHONPATH='.':$PYTHONPATH \
python tools/test.py \
  configs/imp/panoptic_fpn_r50_fpn_1x_sgdet_psg.py \
  path/to/checkpoint.pth \
  --eval sgdet

Submitting for PSG Competition

# sh scripts/imp/submit_panoptic_fpn_r50_sgdet.sh
PYTHONPATH='.':$PYTHONPATH \
python tools/test.py \
  configs/imp/panoptic_fpn_r50_fpn_1x_sgdet_psg.py \
  path/to/checkpoint.pth \
  --submit

OpenPSG: Benchmarking PSG Task

Supported methods (Welcome to Contribute!)

Two-Stage Methods (4)

IMP (CVPR'17)

MOTIFS (CVPR'18)

VCTree (CVPR'19)

GPSNet (CVPR'20)

One-Stage Methods (2)

PSGTR (ECCV'22)

PSGFormer (ECCV'22)

Supported datasets (Welcome to Contribute!)

VG-150 (IJCV'17)
PSG (ECCV'22)

Model Zoo

Method	Backbone	#Epoch	R/mR@20	R/mR@50	R/mR@100	ckpt	SHA256
IMP	ResNet-50	12	16.5 / 6.52	18.2 / 7.05	18.6 / 7.23	link	7be2842b6664e2b9ef6c7c05d27fde521e2401ffe67dbb936438c69e98f9783c
MOTIFS	ResNet-50	12	20.0 / 9.10	21.7 / 9.57	22.0 / 9.69	link	956471959ca89acae45c9533fb9f9a6544e650b8ea18fe62cdead495b38751b8
VCTree	ResNet-50	12	20.6 / 9.70	22.1 / 10.2	22.5 / 10.2	link	e5fdac7e6cc8d9af7ae7027f6d0948bf414a4a605ed5db4d82c5d72de55c9b58
GPSNet	ResNet-50	12	17.8 / 7.03	19.6 / 7.49	20.1 / 7.67	link	98cd7450925eb88fa311a20fce74c96f712e45b7f29857c5cdf9b9dd57f59c51
PSGTR	ResNet-50	60	28.4 / 16.6	34.4 / 20.8	36.3 / 22.1	link	1c4ddcbda74686568b7e6b8145f7f33030407e27e390c37c23206f95c51829ed
PSGFormer	ResNet-50	60	18.0 / 14.8	19.6 / 17.0	20.1 / 17.6	link	2f0015ce67040fa00b65986f6ce457c4f8cc34720f7e47a656b462b696a013b7

Contributing

We appreciate all contributions to improve OpenPSG. We sincerely welcome community users to participate in these projects. Please refer to CONTRIBUTING.md for the contributing guideline.

Acknowledgements

OpenPSG is developed based on MMDetection. Most of the two-stage SGG implementations refer to MMSceneGraph and Scene-Graph-Benchmark.pytorch. We sincerely appreciate the efforts of the developers from the previous codebases.

Citation

If you find our repository useful for your research, please consider citing our paper:

@inproceedings{yang2022psg,
    author = {Yang, Jingkang and Ang, Yi Zhe and Guo, Zujin and Zhou, Kaiyang and Zhang, Wayne and Liu, Ziwei},
    title = {Panoptic Scene Graph Generation},
    booktitle = {ECCV}
    year = {2022}
}

KastanDay/faster_OpenPSG