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 51.5 box AP and 46.0 mask AP on COCO.

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 Table 2 in our paper for a detailed analysis.
• We also provide a training config w/ sample ratio = 0.25 for 8x GPUs (bsz = 16) environment to make our work more accessible to the community. The results (49.9 Box AP / 44.6 Mask AP) match our default settings (49.9 Box AP / 44.7 Mask AP), and are better than the Swin-Base counterpart (49.2 Box AP / 43.5 Mask AP) under a similar total training time (~2d6h).
• Benchmarking-ViT-B is an unofficial implementation of Benchmarking Detection Transfer Learning with Vision Transformers
• The configuration & results of MIMDet-ViT-L are still under-tuned.

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

• 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.

# 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 (see Table 2 in our paper for a 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

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>

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

MIMDet is released under the MIT License.

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}
}