GLIP: Grounded Language-Image Pre-training

Updates

• 01/17/2023: From image understanding to image generation for open-set grounding? Check out GLIGEN (Grounded Language-to-Image Generation)

  • GLIGEN: (box, concept) $\rightarrow$ image || GLIP: image $\rightarrow$ (box, concept)

• 09/19/2022: GLIPv2 has been accepted to NeurIPS 2022 (Updated Version).

• 09/18/2022: Organizing ECCV Workshop Computer Vision in the Wild (CVinW), where two challenges are hosted to evaluate the zero-shot, few-shot and full-shot performance of pre-trained vision models in downstream tasks:

  • ``Image Classification in the Wild (ICinW)'' Challenge evaluates on 20 image classification tasks.
  • ``Object Detection in the Wild (ODinW)'' Challenge evaluates on 35 object detection tasks.

$\qquad$ [Workshop] $\qquad$ [IC Challenge] $\qquad$ [OD Challenge]

• 09/13/2022: Updated HuggingFace Demo! Feel free to give it a try!!!

  • Acknowledgement: Many thanks to the help from @HuggingFace for a Space GPU upgrade to host the GLIP demo!

• 06/21/2022: GLIP has been selected as a Best Paper Finalist at CVPR 2022!

• 06/16/2022: ODinW benchmark released! GLIP-T A&B released!

• 06/13/2022: GLIPv2 is on Arxiv https://arxiv.org/abs/2206.05836!

• 04/30/2022: Updated Colab Demo!

• 04/14/2022: GLIP has been accepted to CVPR 2022 as an oral presentation! First version of code and pre-trained models are released!

• 12/06/2021: GLIP paper on arxiv https://arxiv.org/abs/2112.03857.

• 11/23/2021: Project page built.
  

Introduction

This repository is the project page for GLIP. GLIP demonstrate strong zero-shot and few-shot transferability to various object-level recognition tasks.

1. When directly evaluated on COCO and LVIS (without seeing any images in COCO), GLIP achieves 49.8 AP and 26.9 AP, respectively, surpassing many supervised baselines.
2. After fine-tuned on COCO, GLIP achieves 60.8 AP on val and 61.5 AP on test-dev, surpassing prior SoTA.
3. When transferred to 13 downstream object detection tasks, a few-shot GLIP rivals with a fully-supervised Dynamic Head.

We provide code for:

1. pre-training GLIP on detection and grounding data;
2. zero-shot evaluating GLIP on standard benchmarks (COCO, LVIS, Flickr30K) and custom COCO-formated datasets;
3. fine-tuning GLIP on standard benchmarks (COCO) and custom COCO-formated datasets;
4. a Colab demo.
5. Toolkits for the Object Detection in the Wild Benchmark (ODinW) with 35 downstream detection tasks.

Please see respective sections for instructions.

Demo

Please see a Colab demo at link!

Installation and Setup

Environment This repo requires Pytorch>=1.9 and torchvision. We recommand using docker to setup the environment. You can use this pre-built docker image docker pull pengchuanzhang/maskrcnn:ubuntu18-py3.7-cuda10.2-pytorch1.9 or this one docker pull pengchuanzhang/pytorch:ubuntu20.04_torch1.9-cuda11.3-nccl2.9.9 depending on your GPU.

Then install the following packages:

pip install einops shapely timm yacs tensorboardX ftfy prettytable pymongo
pip install transformers 
python setup.py build develop --user

Backbone Checkpoints. Download the ImageNet pre-trained backbone checkpoints into the MODEL folder.

mkdir MODEL
wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/models/swin_tiny_patch4_window7_224.pth -O swin_tiny_patch4_window7_224.pth
wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/models/swin_large_patch4_window12_384_22k.pth -O swin_large_patch4_window12_384_22k.pth

Model Zoo

Checkpoint host move. The checkpoint links expired. We are moving the checkpoints to https://huggingface.co/harold/GLIP/tree/main. Currently GLIP-T and GLIP-T (C) checkpoints are available. Working to host the remaining checkpoints asap.

Model	COCO [1]	LVIS [2]	LVIS [3]	ODinW [4]	Pre-Train Data	Config	Weight
GLIP-T (A)	42.9 / 52.9	-	14.2	~28.7	O365	config	weight
GLIP-T (B)	44.9 / 53.8	-	13.5	~33.2	O365	config	weight
GLIP-T (C)	46.7 / 55.1	14.3	17.7	44.4	O365,GoldG	config	weight
GLIP-T [5]	46.6 / 55.2	17.6	20.1	42.7	O365,GoldG,CC3M,SBU	config [6]	weight
GLIP-L [7]	51.4 / 61.7 [8]	29.3	30.1	51.2	FourODs,GoldG,CC3M+12M,SBU	config [9]	weight

[1] Zero-shot and fine-tuning performance on COCO val2017.

[2] Zero-shot performance on LVIS minival (APr) with the last pre-trained checkpoint.

[3] On LVIS, the model could overfit slightly during the pre-training course. Thus we reported two numbers on LVIS: the performance of the last checkpoint (LVIS[2]) and the performance of the best checkpoint during the pre-training course (LVIS[3]).

[4] Zero-shot performance on the 13 ODinW datasets. The numbers reported in the GLIP paper is from the best checkpoint during the pre-training course, which may be slightly higher than the numbers for the released last checkpoint, similar to the case of LVIS.

[5] GLIP-T released in this repo is pre-trained on Conceptual Captions 3M and SBU captions. It is referred in paper in Table 1 and in Appendix C.3. It differs slightly from the GLIP-T in the main paper in terms of downstream performance. We will release the pre-training support for using CC3M and SBU captions data in the next update.

[6] This config is only intended for zero-shot evaluation and fine-tuning. Pre-training config with support for using CC3M and SBU captions data will be updated.

[7] GLIP-L released in this repo is pre-trained on Conceptual Captions 3M+12M and SBU captions. It slightly outperforms the GLIP-L in the main paper because the model used to annotate the caption data are improved compared to the main paper. We will release the pre-training support for using CC3M+12M and SBU captions data in the next update.

[8] Multi-scale testing used.

[9] This config is only intended for zero-shot evaluation and fine-tuning. Pre-training config with support for using CC3M+12M and SBU captions data to be updated.

Pre-Training

Required Data. Prepare Objects365, Flickr30K, and MixedGrounding data as in DATA.md. Support for training using caption data (Conceptual Captions and SBU captions) will be released soon.

Command.

Perform pre-training with the following command (please change the config-file accordingly; checkout model zoo for the corresponding config; change the {output_dir} to your desired output directory):

python -m torch.distributed.launch --nnodes 2 --nproc_per_node=16 tools/train_net.py \
    --config-file configs/pretrain/glip_Swin_T_O365_GoldG.yaml \
    --skip-test --use-tensorboard --override_output_dir {output_dir}

For training GLIP-T models, we used nnodes = 2, nproc_per_node=16 on 32GB V100 machines. For training GLIP-L models, we used nnodes = 4, nproc_per_node=16 on 32GB V100 machines. Please setup the environment accordingly based on your local machine.

(Zero-Shot) Evaluation

COCO Evaluation

Prepare COCO/val2017 data as in DATA.md. Set {config_file}, {model_checkpoint} according to the Model Zoo; set {output_dir} to a folder where the evaluation results will be stored.

python tools/test_grounding_net.py --config-file {config_file} --weight {model_checkpoint} \
        TEST.IMS_PER_BATCH 1 \
        MODEL.DYHEAD.SCORE_AGG "MEAN" \
        TEST.EVAL_TASK detection \
        MODEL.DYHEAD.FUSE_CONFIG.MLM_LOSS False \
        OUTPUT_DIR {output_dir}

LVIS Evaluation

We follow MDETR to evaluate with the FixedAP criterion. Set {config_file}, {model_checkpoint} according to the Model Zoo. Prepare COCO/val2017 data as in DATA.md.

python -m torch.distributed.launch --nproc_per_node=4 \
        tools/test_grounding_net.py \
        --config-file {config_file} \
        --task_config configs/lvis/minival.yaml \
        --weight {model_checkpoint} \
        TEST.EVAL_TASK detection OUTPUT_DIR {output_dir} 
        TEST.CHUNKED_EVALUATION 40  TEST.IMS_PER_BATCH 4 SOLVER.IMS_PER_BATCH 4 TEST.MDETR_STYLE_AGGREGATE_CLASS_NUM 3000 MODEL.RETINANET.DETECTIONS_PER_IMG 300 MODEL.FCOS.DETECTIONS_PER_IMG 300 MODEL.ATSS.DETECTIONS_PER_IMG 300 MODEL.ROI_HEADS.DETECTIONS_PER_IMG 300

If you wish to evaluate on Val 1.0, set --task_config to configs/lvis/val.yaml.

ODinW / Custom Dataset Evaluation

GLIP supports easy evaluation on a custom dataset. Currently, the code supports evaluation on COCO-formatted dataset.

We will use the Aquarium dataset from ODinW as an example to show how to evaluate on a custom COCO-formatted dataset.

1. Download the raw dataset from RoboFlow in the COCO format into DATASET/odinw/Aquarium. Each train/val/test split has a corresponding annotation file and a image folder.

2. Remove the background class from the annotation file. This can be as simple as open "_annotations.coco.json" and remove the entry with "id:0" from "categories". For convenience, we provide the modified annotation files for Aquarium:

   wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/odinw/Aquarium/Aquarium%20Combined.v2-raw-1024.coco/test/annotations_without_background.json -O DATASET/odinw/Aquarium/Aquarium\ Combined.v2-raw-1024.coco/test/annotations_without_background.json
   wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/odinw/Aquarium/Aquarium%20Combined.v2-raw-1024.coco/train/annotations_without_background.json -O DATASET/odinw/Aquarium/Aquarium\ Combined.v2-raw-1024.coco/train/annotations_without_background.json
   wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/odinw/Aquarium/Aquarium%20Combined.v2-raw-1024.coco/valid/annotations_without_background.json -O DATASET/odinw/Aquarium/Aquarium\ Combined.v2-raw-1024.coco/valid/annotations_without_background.json
   

3. Then create a yaml file as in configs/odinw_13/Aquarium_Aquarium_Combined.v2-raw-1024.coco.yaml. A few fields to be noted in the yamls:

   DATASET.CAPTION_PROMPT allows manually changing the prompt (the default prompt is simply concatnating all the categories);

   MODELS.*.NUM_CLASSES need to be set to the number of categories in the dataset (including the background class). E.g., Aquarium has 7 non-background categories thus MODELS.*.NUM_CLASSES is set to 8;

4. Run the following command to evaluate on the dataset. Set {config_file}, {model_checkpoint} according to the Model Zoo. Set {odinw_configs} to the path of the task yaml file we just prepared.

python tools/test_grounding_net.py --config-file {config_file} --weight {model_checkpoint} \
      --task_config {odinw_configs} \
      TEST.IMS_PER_BATCH 1 SOLVER.IMS_PER_BATCH 1 \
      TEST.EVAL_TASK detection \
      DATASETS.TRAIN_DATASETNAME_SUFFIX _grounding \
      DATALOADER.DISTRIBUTE_CHUNK_AMONG_NODE False \
      DATASETS.USE_OVERRIDE_CATEGORY True \
      DATASETS.USE_CAPTION_PROMPT True

Flickr30K Evaluation

Prepare Flickr30K data as in DATA.md. Set {config_file}, {model_checkpoint} according to the Model Zoo.

python tools/test_grounding_net.py \
        --config-file {config_file} \
        --task_config configs/flickr/test.yaml,configs/flickr/val.yaml \
        --weight {model_checkpoint} \
        OUTPUT_DIR {output_dir} TEST.IMS_PER_BATCH 1 SOLVER.IMS_PER_BATCH 1 TEST.MDETR_STYLE_AGGREGATE_CLASS_NUM 100 TEST.EVAL_TASK grounding MODEL.DYHEAD.FUSE_CONFIG.MLM_LOSS False

Fine-Tuning

COCO Fine-Tuning

Prepare the COCO data as in DATA.md. Set {config_file}, {model_checkpoint} according to the Model Zoo.

Below is the fine-tuning script for tuning the Tiny models:

python -m torch.distributed.launch --nproc_per_node=16 tools/train_net.py \
       --config-file {config_file} \
       --skip-test \
       MODEL.WEIGHT {model_checkpoint} \
       DATASETS.TRAIN '("coco_grounding_train", )' \
       MODEL.BACKBONE.FREEZE_CONV_BODY_AT -1 SOLVER.IMS_PER_BATCH 32 SOLVER.USE_AMP True SOLVER.MAX_EPOCH 24 TEST.DURING_TRAINING False TEST.IMS_PER_BATCH 16 SOLVER.FIND_UNUSED_PARAMETERS False SOLVER.BASE_LR 0.00001 SOLVER.LANG_LR 0.00001 SOLVER.STEPS \(0.67,0.89\) DATASETS.DISABLE_SHUFFLE True MODEL.DYHEAD.SCORE_AGG "MEAN" TEST.EVAL_TASK detection

For evaluation, please follow the instructions in COCO Evaluation. Scripts for tuning the Large model will be released soon.

ODinW / Custom Dataset Fine-Tuning

Prepare the dataset as in ODinW / Custom Dataset Evaluation.

Full Model Fine-Tuning

For tuning with 1/3/5/10-shot, set {custom_shot_and_epoch_and_general_copy} to "1_200_8", "3_200_4", "5_200_2", "10_200_1", respectively.

For tuning with all the data, set {custom_shot_and_epoch_and_general_copy} to "0_200_1"; set SOLVER.STEP_PATIENCE to 2; set SOLVER.AUTO_TERMINATE_PATIENCE to 4.

python -m torch.distributed.launch --nproc_per_node=4 tools/finetune.py \
      --config-file {config_file}  --ft-tasks {configs} --skip-test \
      --custom_shot_and_epoch_and_general_copy {custom_shot_and_epoch_and_general_copy} \
      --evaluate_only_best_on_test --push_both_val_and_test \
      MODEL.WEIGHT {model_checkpoint} \
      SOLVER.USE_AMP True TEST.DURING_TRAINING True TEST.IMS_PER_BATCH 4 SOLVER.IMS_PER_BATCH 4 SOLVER.WEIGHT_DECAY 0.05 TEST.EVAL_TASK detection DATASETS.TRAIN_DATASETNAME_SUFFIX _grounding MODEL.BACKBONE.FREEZE_CONV_BODY_AT 2 MODEL.DYHEAD.USE_CHECKPOINT True SOLVER.FIND_UNUSED_PARAMETERS False SOLVER.TEST_WITH_INFERENCE True SOLVER.USE_AUTOSTEP True DATASETS.USE_OVERRIDE_CATEGORY True SOLVER.SEED 10 DATASETS.SHUFFLE_SEED 3 DATASETS.USE_CAPTION_PROMPT True DATASETS.DISABLE_SHUFFLE True \
      SOLVER.STEP_PATIENCE 3 SOLVER.CHECKPOINT_PER_EPOCH 1.0 SOLVER.AUTO_TERMINATE_PATIENCE 8 SOLVER.MODEL_EMA 0.0 SOLVER.TUNING_HIGHLEVEL_OVERRIDE full

Prompt Tuning

Follow the command as in Full Model Fine-Tuning. But set the following hyper-parameters:

SOLVER.WEIGHT_DECAY 0.25 \
SOLVER.BASE_LR 0.05 \
SOLVER.TUNING_HIGHLEVEL_OVERRIDE language_prompt_v2

The Object Detection in the Wild Benchmark

ODinW was first proposed in GLIP and refined and formalized in ELEVATER. GLIP used 13 downstream tasks while the full ODinW has 35 downstream tasks. It will be hosted as a challenge at the CV in the Wild Workshop @ ECCV 2022. We hope our code encourage the community to participate in this challenge!

ODinW was introduced in GLIP and initially contained 13 datasets. We further expand the datasets by including more datasets from RoboFlow and the final version contains 35 datasets.

To distinguish between the two versions, we denote the version used by GLIP as ODinW-13 and the version used by the CVinW workshop as ODinW-35.

This repo also provides the necessary code to train and evaluate on ODinW. See instructions below.

Download ODinW

RoboFlow hosts all the original datasets. We are also hosting the datasets and provide a simple script the download all the data.

python odinw/download_datasets.py

configs/odinw_35 contain all the meta information of the datasets. configs/odinw_13 are the datasets used by GLIP. Each dataset follows the coco detection format.

All ODinW datasets are in the COCO format; thus we can directly use the similar scripts to adapt and evaluate pre-trained models on ODinW. Below is a brief recap.

(Zero-Shot) Evaluation

odinw_configs can be any of the configs from configs/odinw_14 and configs/odinw_35.

python tools/test_grounding_net.py --config-file {config_file} --weight {model_checkpoint} \
      --task_config {odinw_configs} \
      TEST.IMS_PER_BATCH 1 SOLVER.IMS_PER_BATCH 1 \
      TEST.EVAL_TASK detection \
      DATASETS.TRAIN_DATASETNAME_SUFFIX _grounding \
      DATALOADER.DISTRIBUTE_CHUNK_AMONG_NODE False \
      DATASETS.USE_OVERRIDE_CATEGORY True \
      DATASETS.USE_CAPTION_PROMPT True

Full-Model Fine-Tuning

For tuning with 1/3/5/10-shot, set {custom_shot_and_epoch_and_general_copy} to "1_200_8", "3_200_4", "5_200_2", "10_200_1", respectively.

For tuning with all the data, set {custom_shot_and_epoch_and_general_copy} to "0_200_1"; set SOLVER.STEP_PATIENCE to 2; set SOLVER.AUTO_TERMINATE_PATIENCE to 4.

python -m torch.distributed.launch --nproc_per_node=4 tools/finetune.py \
      --config-file {config_file}  --ft-tasks {odinw_configs} --skip-test \
      --custom_shot_and_epoch_and_general_copy {custom_shot_and_epoch_and_general_copy} \
      --evaluate_only_best_on_test --push_both_val_and_test \
      MODEL.WEIGHT {model_checkpoint} \
      SOLVER.USE_AMP True TEST.DURING_TRAINING True TEST.IMS_PER_BATCH 4 SOLVER.IMS_PER_BATCH 4 SOLVER.WEIGHT_DECAY 0.05 TEST.EVAL_TASK detection DATASETS.TRAIN_DATASETNAME_SUFFIX _grounding MODEL.BACKBONE.FREEZE_CONV_BODY_AT 2 MODEL.DYHEAD.USE_CHECKPOINT True SOLVER.FIND_UNUSED_PARAMETERS False SOLVER.TEST_WITH_INFERENCE True SOLVER.USE_AUTOSTEP True DATASETS.USE_OVERRIDE_CATEGORY True SOLVER.SEED 10 DATASETS.SHUFFLE_SEED 3 DATASETS.USE_CAPTION_PROMPT True DATASETS.DISABLE_SHUFFLE True \
      SOLVER.STEP_PATIENCE 3 SOLVER.CHECKPOINT_PER_EPOCH 1.0 SOLVER.AUTO_TERMINATE_PATIENCE 8 SOLVER.MODEL_EMA 0.0 SOLVER.TUNING_HIGHLEVEL_OVERRIDE full

Prompt Tuning

For tuning with 1/3/5/10-shot, set {custom_shot_and_epoch_and_general_copy} to "1_200_8", "3_200_4", "5_200_2", "10_200_1", respectively.

For tuning with all the data, set {custom_shot_and_epoch_and_general_copy} to "0_200_1"; set SOLVER.STEP_PATIENCE to 2; set SOLVER.AUTO_TERMINATE_PATIENCE to 4.

Follow the command as in Full Model Fine-Tuning. But set the following hyper-parameters:

SOLVER.WEIGHT_DECAY 0.25 \
SOLVER.BASE_LR 0.05 \
SOLVER.TUNING_HIGHLEVEL_OVERRIDE language_prompt_v2

Linear Probing

For tuning with 1/3/5/10-shot, set {custom_shot_and_epoch_and_general_copy} to "1_200_8", "3_200_4", "5_200_2", "10_200_1", respectively.

For tuning with all the data, set {custom_shot_and_epoch_and_general_copy} to "0_200_1"; set SOLVER.STEP_PATIENCE to 2; set SOLVER.AUTO_TERMINATE_PATIENCE to 4.

Follow the command as in Full Model Fine-Tuning. But set the following hyper-parameters:

SOLVER.TUNING_HIGHLEVEL_OVERRIDE linear_prob

Knowledge-Augmented Inference

GLIP also supports knowledge-augmented inference. Please see our paper for details. Here we provide an example on how to use external knowledge. Please download a specialized GLIP-A model for knowledge augmented inference wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/models/glip_a_tiny_o365_knowledge.pth -O MODEL/glip_a_tiny_o365_knowledge.pth.

python tools/test_grounding_net.py --config-file configs/pretrain/glip_A_Swin_T_O365.yaml --weight MODEL/glip_a_tiny_o365_knowledge.pth \
      --task_config {odinw_configs} \
      TEST.IMS_PER_BATCH 1 SOLVER.IMS_PER_BATCH 1 \
      TEST.EVAL_TASK detection \
      DATASETS.TRAIN_DATASETNAME_SUFFIX _grounding \
      DATALOADER.DISTRIBUTE_CHUNK_AMONG_NODE False \
      DATASETS.USE_OVERRIDE_CATEGORY True \
      DATASETS.USE_CAPTION_PROMPT True \
      GLIPKNOW.KNOWLEDGE_FILE knowledge/odinw_benchmark35_knowledge_and_gpt3.yaml GLIPKNOW.KNOWLEDGE_TYPE gpt3_and_wiki GLIPKNOW.PARALLEL_LANGUAGE_INPUT True GLIPKNOW.LAN_FEATURE_AGG_TYPE first MODEL.DYHEAD.FUSE_CONFIG.USE_LAYER_SCALE True GLIPKNOW.GPT3_NUM 3 GLIPKNOW.WIKI_AND_GPT3 True

Submit Your Results to ODinw Leaderboard

The participant teams are encouraged to upload their results to ODinW leaderboard on EvalAI. From the perspective od data labeling cost, lowering the requirement of data requirement enables more scenarios, a varied number of tracks are considered in the challenge: zero-shot, few-shot, and full-shot. Please see the ODinW website for more details about each phase.

1. For zero/full shot setting, the required format for prediction json file is

{
      "dataset_name (e.g., 'WildFireSmoke')":
            [value]: value is following the COCO's 
            result format, which contains 
            ["image_id":xxx, "category_id":xxx, 
            "bbox":xxx, "score":xxx]
}

Please see one provided example for zero shot prediction file: all_predictions_zeroshot.json and one full shot prediction file: all_predictions_fullshot.json.

2. For few shot (3-shot, according to the challenge description) setting, where three train-val subsets are generated with random seed [3, 30, 300], respectively. The required format for prediction json file is

{
      "dataset_name (e.g., "WildFireSmoke")":{
            "rand_seed_num (e.g., "30")":
                  [value]: value is following the 
                  COCO's result format, which 
                  contains ["image_id":xxx, 
                  "category_id":xxx, "bbox":xxx, 
                  "score":xxx]
     }
}

Please see one provided example for few shot prediction file: all_predictions_3_shot.json.

Citations

Please consider citing our papers if you use the code:

@inproceedings{li2021grounded,
      title={Grounded Language-Image Pre-training},
      author={Liunian Harold Li* and Pengchuan Zhang* and Haotian Zhang* and Jianwei Yang and Chunyuan Li and Yiwu Zhong and Lijuan Wang and Lu Yuan and Lei Zhang and Jenq-Neng Hwang and Kai-Wei Chang and Jianfeng Gao},
      year={2022},
      booktitle={CVPR},
}
@article{zhang2022glipv2,
  title={GLIPv2: Unifying Localization and Vision-Language Understanding},
  author={Zhang, Haotian* and Zhang, Pengchuan* and Hu, Xiaowei and Chen, Yen-Chun and Li, Liunian Harold and Dai, Xiyang and Wang, Lijuan and Yuan, Lu and Hwang, Jenq-Neng and Gao, Jianfeng},
  journal={arXiv preprint arXiv:2206.05836},
  year={2022}
}
@article{li2022elevater,
  title={ELEVATER: A Benchmark and Toolkit for Evaluating Language-Augmented Visual Models},
  author={Li*, Chunyuan and Liu*, Haotian and Li, Liunian Harold and Zhang, Pengchuan and Aneja, Jyoti and Yang, Jianwei and Jin, Ping and Lee, Yong Jae and Hu, Houdong and Liu, Zicheng and others},
  journal={arXiv preprint arXiv:2204.08790},
  year={2022}
}