ConvMAE: Masked Convolution Meets Masked Autoencoders

[arXiv]

This repository contains the implementation of the ConvMAE transfer learning for object detection on COCO.

For ImageNet pretraining and pretrained checkpoint, please refer to ConvMAE.

COCO

Name	pre-train	pre-train epoch	finetune epoch	box AP	mask AP	model	log
ViTDet, ViT-B	IN1K, MAE	1600	100	51.6	45.9	-	-
ViTDet, ConvMAE-B	IN1K, ConvMAE	1600	25	53.9	47.6	model	log

Installation

Please follow Installation to install detectron2.

Preparing Dataset

cd datasets
ln -s /path/to/coco coco

Training

python tools/lazyconfig_train_net.py --num-gpus 8 --config-file \ 
projects/ConvMAEDet/configs/COCO/mask_rcnn_vitdet_convmae_b_25ep.py \
train.init_checkpoint=path/to/pretrained_model

Evaluation

python tools/lazyconfig_train_net.py --num-gpus 8 --eval-only --config-file \ 
projects/ConvMAEDet/configs/COCO/mask_rcnn_vitdet_convmae_b_25ep.py \
train.init_checkpoint=path/to/model_checkpoint

Acknowledgement

This project is based on Detectron2 and VitDet. Thanks for their wonderful work.

Citation

@article{gao2022convmae,
  title={ConvMAE: Masked Convolution Meets Masked Autoencoders},
  author={Gao, Peng and Ma, Teli and Li, Hongsheng and Dai, Jifeng and Qiao, Yu},
  journal={arXiv preprint arXiv:2205.03892},
  year={2022}
}