Feng Li*, Hao Zhang*, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M. Ni, and Heung-Yeung Shum.

This repository is the official implementation of the Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation (DINO pronounced `daɪnoʊ' as in dinosaur). Our code is based on detectron2. detrex version is opensource simultaneously.

🔥 We release a strong open-set object detection and segmentation model OpenSeeD based on MaskDINO that achieves the best results on open-set object segmentation tasks. Code and checkpoints are available here.

• A unified architecture for object detection, panoptic, instance and semantic segmentation.
• Achieve task and data cooperation between detection and segmentation.
• State-of-the-art performance under the same setting.
• Support major detection and segmentation datasets: COCO, ADE20K, Cityscapes.

• [2022/12/02] Our code and checkpoints are available! Mask DINO further Achieves 51.7 and 59.0 box AP on COCO with a ResNet-50 and SwinL without extra detection data, outperforming DINO under the same setting!

• [2022/6] We propose a unified detection and segmentation model Mask DINO that achieves the best results on all the three segmentation tasks (54.7 AP on COCO instance leaderboard, 59.5 PQ on COCO panoptic leaderboard, and 60.8 mIoU on ADE20K semantic leaderboard)!

See installation instructions.

See Inference Demo with Pre-trained Model

See Results.

See Preparing Datasets for MaskDINO.

See Getting Started.

See More Usage.

In this part, we present the clean models that do not use extra detection data or tricks.

we follow DINO to use hidden dimension 2048 in the encoder of feedforward by default. We also use the mask-enhanced box initialization proposed in our paper in instance segmentation and detection. To better present our model, we also list the models trained with hidden dimension 1024 (hid 1024) and not using mask-enhance initialization (no mask enhance) in this table.

We use hidden dimension 1024 and 100 queries for semantic segmentation.

You can also find all these models here.

All models were trained with 4 NVIDIA A100 GPUs (ResNet-50 based models) or 8 NVIDIA A100 GPUs (Swin-L based models).

We will release more pretrained models in the future.

In the above tables, the "Name" column contains a link config_path to the config file, and the corresponding model checkpoints can be downloaded from the link in model.

If your dataset files are not under this repo, you need to add export DETECTRON2_DATASETS=/path/to/your/data or use Symbolic Link ln -s to link the dataset into this repo before the following command first.

• You can download our pretrained models and evaluate them with the following commands.

  python train_net.py --eval-only --num-gpus 8 --config-file config_path MODEL.WEIGHTS /path/to/checkpoint_file

  for example, to reproduce our instance segmentation result, you can copy the config path from the table, download the pretrained checkpoint into /path/to/checkpoint_file, and run

  python train_net.py --eval-only --num-gpus 8 --config-file configs/coco/instance-segmentation/maskdino_R50_bs16_50ep_3s_dowsample1_2048.yaml MODEL.WEIGHTS /path/to/checkpoint_file

  which can reproduce the model.

• Use the above command without eval-only will train the model. For Swin backbones, you need to specify the path of the pretrained backbones with MODEL.WEIGHTS /path/to/pretrained_checkpoint

  python train_net.py --num-gpus 8 --config-file config_path MODEL.WEIGHTS /path/to/checkpoint_file

• For ResNet-50 models, training on 8 GPU requires around 15G memory on each GPU and 3 days training for 50 epochs.
• For Swin-L models, training on 8 gpu required memory 60G on each GPU. If your gpu do not have enough memory, you may also train with 16 GPUs with distributed training on two nodes.
• We use total batch size 16 for all our models. If train on 1 GPU, you need to figure out learning rate and batch size by yourself

  python train_net.py --num-gpus 1 --config-file config_path SOLVER.IMS_PER_BATCH SET_TO_SOME_REASONABLE_VALUE SOLVER.BASE_LR SET_TO_SOME_REASONABLE_VALUE

You can also refer to Getting Started with Detectron2 for full usage.

We provide 2 ways to convert predicted masks to boxes to initialize decoder boxes. You can set as follows

• MODEL.MaskDINO.INITIALIZE_BOX_TYPE: no not using mask enhanced box initialization
• MODEL.MaskDINO.INITIALIZE_BOX_TYPE: mask2box a fast conversion way
• MODEL.MaskDINO.INITIALIZE_BOX_TYPE: bitmask provided conversion from detectron2, slower but more accurate conversion.

These two conversion ways do not affect the final performance much, you can choose either way.

In addition, if you already train a model for 50 epochs without mask-enhance box initialization, you can plug in this method and simply finetune the model in the last few epochs (i.e., load from 32K iteration trained model and finetune it). This way can also achieve similar performance compared with training from scratch, but more flexible.

MaskDINO consists of three components: a backbone, a pixel decoder and a Transformer decoder. You can easily replace each of these three components with your own implementation.

• backbone: Define and register your backbone under maskdino/modeling/backbone. You can follow the Swin Transformer as an example.

• pixel decoder: pixel decoder is actually the multi-scale encoder in DINO and Deformable DETR, we follow mask2former to call it pixel decoder. It is in maskdino/modeling/pixel_decoder, you can change your multi-scale encoder. The returned values include

  1. mask_features is the per-pixel embeddings with resolution 1/4 of the original image, obtained by fusing backbone 1/4 features and multi-scale encoder encoded 1/8 features. This is used to produce binary masks.
  2. multi_scale_features, which is the multi-scale inputs to the Transformer decoder. For ResNet-50 models with 4 scales, we use resolution 1/32, 1/16, and 1/8 but you can use arbitrary resolutions here, and follow DINO to additionally downsample 1/32 to get a 4th scale with 1/64 resolution. For 5-scale models with SwinL, we additional use 1/4 resolution features as in DINO.

• transformer decoder: it mainly follows DINO decoder to do detection and segmentation tasks. It is defined in maskdino/modeling/transformer_decoder.

Mask DINO is released under the Apache 2.0 license. Please see the LICENSE file for more information.

Copyright (c) IDEA. All rights reserved.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use these files except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

If you find our work helpful for your research, please consider citing the following BibTeX entry.

@misc{li2022mask,
      title={Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation}, 
      author={Feng Li and Hao Zhang and Huaizhe xu and Shilong Liu and Lei Zhang and Lionel M. Ni and Heung-Yeung Shum},
      year={2022},
      eprint={2206.02777},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

If you find the code useful, please also consider the following BibTeX entry.

@misc{zhang2022dino,
      title={DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection}, 
      author={Hao Zhang and Feng Li and Shilong Liu and Lei Zhang and Hang Su and Jun Zhu and Lionel M. Ni and Heung-Yeung Shum},
      year={2022},
      eprint={2203.03605},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@inproceedings{li2022dn,
      title={Dn-detr: Accelerate detr training by introducing query denoising},
      author={Li, Feng and Zhang, Hao and Liu, Shilong and Guo, Jian and Ni, Lionel M and Zhang, Lei},
      booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
      pages={13619--13627},
      year={2022}
}

@inproceedings{
      liu2022dabdetr,
      title={{DAB}-{DETR}: Dynamic Anchor Boxes are Better Queries for {DETR}},
      author={Shilong Liu and Feng Li and Hao Zhang and Xiao Yang and Xianbiao Qi and Hang Su and Jun Zhu and Lei Zhang},
      booktitle={International Conference on Learning Representations},
      year={2022},
      url={https://openreview.net/forum?id=oMI9PjOb9Jl}
}

Many thanks to these excellent opensource projects

• Mask2Former
• DINO