News

• 19 May, 2022: We update our preprint with stronger results and more analysis. Code & models are also updated in the main branch. For our previous results (code & models), please refer to the v1.0.0 branch.

• 6 Apr, 2022: Code & models are released!

Introduction

This repo provides code and pretrained models for MIMDet (Masked Image Modeling for Detection).

• MIMDet is a simple framekwork that enables a MIM pretrained vanilla ViT to perform high-performance object-level understanding, e.g, object detection and instance segmentation.
• In MIMDet, a MIM pre-trained vanilla ViT encoder can work surprisingly well in the challenging object-level recognition scenario even with randomly sampled partial observations, e.g., only 25%~50% of the input embeddings.
• In order to construct multi-scale representations for object detection, a randomly initialized compact convolutional stem supplants the pre-trained large kernel patchify stem, and its intermediate features can naturally serve as the higher resolution inputs of a feature pyramid without upsampling. While the pre-trained ViT is only regarded as the third-stage of our detector's backbone instead of the whole feature extractor, resulting in a ConvNet-ViT hybrid architecture.
• MIMDet w/ ViT-Base & Mask R-CNN FPN obtains 51.7 box AP and 46.2 mask AP on COCO. With ViT-L, MIMDet achieves 54.3 box AP and 48.2 mask AP.
• We also provide an unofficial implementation of Benchmarking Detection Transfer Learning with Vision Transformers that successfully reproduces its reported results.

Models and Main Results

Mask R-CNN

Notes:

• The Box AP & Mask AP in the table above is obtained w/ sample ratio = 1.0, which is higher than the training sample ratio (0.25 or 0.5). Our MIMDet can benefit from lower sample ratio during training for better efficiency, as well as higher sample ratio during inference for better accuracy. Please refer to our paper for detailed analysis.
• Benchmarking-ViT-B is an unofficial implementation of Benchmarking Detection Transfer Learning with Vision Transformers.

Installation

Prerequisites

• Linux
• Python 3.7+
• CUDA 10.2+
• GCC 5+

Prepare

• Clone

git clone https://github.com/hustvl/MIMDet.git
cd MIMDet

• Create a conda virtual environment and activate it:

conda create -n mimdet python=3.9
conda activate mimdet

• Install torch==1.9.0 and torchvision==0.10.0
• Install Detectron2==0.6, follow d2 doc.
• Install timm==0.4.12, follow timm doc.
• Install einops, follow einops repo.
• Prepare COCO dataset, follow d2 doc.

Dataset

MIMDet is built upon detectron2, so please organize dataset directory in detectron2's manner. We refer users to detectron2 for detailed instructions. The overall hierachical structure is illustrated as following:

MIMDet
├── datasets
│   ├── coco
│   │   ├── annotations
│   │   ├── train2017
│   │   ├── val2017
│   │   ├── test2017
│   ├── ...
├── ...

Training

Download the full MAE pretrained (including the decoder) ViT-B Model and ViT-L Model checkpoint. See MAE repo-issues-8.

# single-machine training
python lazyconfig_train_net.py --config-file <CONFIG_FILE> --num-gpus <GPU_NUM> mae_checkpoint.path=<MAE_MODEL_PATH>

# multi-machine training
python lazyconfig_train_net.py --config-file <CONFIG_FILE> --num-gpus <GPU_NUM> --num-machines <MACHINE_NUM> --master_addr <MASTER_ADDR> --master_port <MASTER_PORT> mae_checkpoint.path=<MAE_MODEL_PATH>

Inference

# inference
python lazyconfig_train_net.py --config-file <CONFIG_FILE> --num-gpus <GPU_NUM> --eval-only train.init_checkpoint=<MODEL_PATH>

# inference with 100% sample ratio (please refer to our paper for detailed analysis)
python lazyconfig_train_net.py --config-file <CONFIG_FILE> --num-gpus <GPU_NUM> --eval-only train.init_checkpoint=<MODEL_PATH> model.backbone.bottom_up.sample_ratio=1.0

Acknowledgement

This project is based on MAE, Detectron2 and timm. Thanks for their wonderful works.

License

MIMDet is released under the MIT License.

Citation

If you find our paper and code useful in your research, please consider giving a star ⭐ and citation 📝 :)

@article{MIMDet,
  title={Unleashing Vanilla Vision Transformer with Masked Image Modeling for Object Detection},
  author={Fang, Yuxin and Yang, Shusheng and Wang, Shijie and Ge, Yixiao and Shan, Ying and Wang, Xinggang},
  journal={arXiv preprint arXiv:2204.02964},
  year={2022}
}