MDQE: Mining Discriminative Query Embeddings to Segment Occluded Instances on Challenging Videos

Minghan LI, Shuai LI, Wangmeng XAING, Lei ZHANG

[arXiv]

Updates

• March 31, 2023: Trained models are released.
• March 28, 2023: Code and paper are now available!

Installation

See installation instructions.

Getting Started

We provide a script train_net.py, that is made to train all the configs provided in MDQE.

Before training: To train a model with "train_net.py" on VIS, first setup the corresponding datasets following Preparing Datasets for MDQE.

Then download pretrained weights in the Model Zoo into the path 'pretrained/coco/*.pth', and run:

python train_net.py --num-gpus 8 \
  --config-file configs/R50_ovis_360.yaml 

To evaluate a model's performance, use

python train_net.py \
  --config-file configs/R50_ovis_360.yaml \
  --eval-only \
  MODEL.WEIGHTS /path/to/checkpoint_file

Model Zoo

Pretrained weights on COCO

Name	R50	Swin-L
MDQE	model, config	model, config

OVIS

Name	Backbone	Frames	AP	Download
MDQE	R50	f4+360p	30.7	model, config
MDQE	R50	f4+640p	32.3	model, config
MDQE	Swin-L	f2+480p	41.0	model, config
MDQE	Swin-L	f2+640p	42.6	model, config

YouTubeVIS-2021

Name	Backbone	Frames	AP	Download
MDQE	R50	f4+360p	46.6	model, config
MDQE	Swin-L	f3+360p	55.5	model, config

YouTubeVIS-2019

Name	Backbone	Frames	AP	Download
MDQE	R50	f4+360p	47.8	model, config
MDQE	Swin-L	f3+360p	59.9	model, config

License

The majority of MDQE is licensed under the Apache-2.0 License. However, portions of the project are available under separate license terms: Detectron2(Apache-2.0 License), IFC(Apache-2.0 License), VITA(Apache-2.0 License), and Deformable-DETR(Apache-2.0 License).

Citing MDQE

If you use MDQE in your research or wish to refer to the baseline results published in the Model Zoo, please use the following BibTeX entry.

@misc{li2023mdqe,
    title={MDQE: Mining Discriminative Query Embeddings to Segment Occluded Instances on Challenging Videos},
    author={Minghan Li and Shuai Li and Wangmeng Xiang and Lei Zhang},
    year={2023},
    eprint={2303.14395},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Acknowledgement

Our code is largely based on Detectron2, IFC, Deformable DETR and VITA. We are truly grateful for their excellent work.