Revealing the Dark Secrets of Masked Image Modeling (Depth Estimation) [Paper]

Main results

Results on NYUv2

Results on KITTI

Preparation

Please refer to [GLPDepth] for configuring the environment and preparing the NYUV2 and KITTI datasets. You can download pretrained models and our well-trained models from zoo(OneDrive).

Training

• Training with model (NYU Depth V2 Swin-Base)

  $ python3 train.py --dataset nyudepthv2 --data_path ../data/ --max_depth 10.0 --max_depth_eval 10.0  --backbone swin_base_v2 --depths 2 2 18 2 --num_filters 32 32 32 --deconv_kernels 2 2 2 --window_size 30 30 30 15 --pretrain_window_size 12 12 12 6 --use_shift True True False False --flip_test --shift_window_test --shift_size 2 --pretrained weights/swin_v2_base_simmim.pth --save_model --crop_h 480 --crop_w 480 --layer_decay 0.9 --drop_path_rate 0.3 --log_dir logs/ 
  

• Training with model (NYU Depth V2 Swin-Large)

  $ python3 train.py --dataset nyudepthv2 --data_path ../data/ --max_depth 10.0 --max_depth_eval 10.0  --backbone swin_large_v2 --depths 2 2 18 2 --num_filters 32 32 32 --deconv_kernels 2 2 2 --window_size 30 30 30 15 --pretrain_window_size 12 12 12 6 --use_shift True True False False --flip_test --shift_window_test --shift_size 2 --pretrained weights/swin_v2_large_simmim.pth --save_model --crop_h 480 --crop_w 480 --layer_decay 0.85 --drop_path_rate 0.5 --log_dir logs/ 
  

• Training with model (KITTI Swin-Base)

  $ python3 train.py --dataset kitti --kitti_crop garg_crop --data_path ../data/ --max_depth 80.0 --max_depth_eval 80.0 --backbone swin_base_v2 --depths 2 2 18 2 --num_filters 32 32 32 --deconv_kernels 2 2 2 --window_size 22 22 22 11 --pretrain_window_size 12 12 12 6 --use_shift True True False False --flip_test --shift_window_test --shift_size 16 --pretrained weights/swin_v2_base_simmim.pth --save_model --crop_h 352 --crop_w 352 --layer_decay 0.9 --drop_path_rate 0.3 --log_dir logs/ 
  

• Training with model (KITTI Swin-Large)

  $ python3 train.py --dataset kitti --kitti_crop garg_crop --data_path ../data/ --max_depth 80.0 --max_depth_eval 80.0 --backbone swin_large_v2 --depths 2 2 18 2 --num_filters 32 32 32 --deconv_kernels 2 2 2 --window_size 22 22 22 11 --pretrain_window_size 12 12 12 6 --use_shift True True False False --flip_test --shift_window_test --shift_size 16 --pretrained weights/swin_v2_large_simmim.pth --save_model --crop_h 352 --crop_w 352 --layer_decay 0.85 --drop_path_rate 0.5 --log_dir logs/ 
  

Evaluation

• Evaluate with model (NYU Depth V2 Swin-Base)

  $ python3 test.py --dataset nyudepthv2 --data_path ../data/ --max_depth 10.0 --max_depth_eval 10.0  --backbone swin_base_v2 --depths 2 2 18 2 --num_filters 32 32 32 --deconv_kernels 2 2 2 --window_size 30 30 30 15 --pretrain_window_size 12 12 12 6 --use_shift True True False False --flip_test --shift_window_test --shift_size 2 --do_evaluate --ckpt_dir ckpt/nyudepthv2_swin_base.ckpt
  

• Evaluate with model (NYU Depth V2 Swin-Large)

  $ python3 test.py --dataset nyudepthv2 --data_path ../data/ --max_depth 10.0 --max_depth_eval 10.0  --backbone swin_large_v2 --depths 2 2 18 2 --num_filters 32 32 32 --deconv_kernels 2 2 2 --window_size 30 30 30 15 --pretrain_window_size 12 12 12 6 --use_shift True True False False --flip_test --shift_window_test --shift_size 2 --do_evaluate --ckpt_dir ckpt/nyudepthv2_swin_large.ckpt
  

• Evaluate with model (KITTI Swin-Base)

  $ python3 test.py --dataset kitti --kitti_crop garg_crop --data_path ../data/ --max_depth 80.0 --max_depth_eval 80.0 --backbone swin_base_v2 --depths 2 2 18 2 --num_filters 32 32 32 --deconv_kernels 2 2 2 --window_size 22 22 22 11 --pretrain_window_size 12 12 12 6 --use_shift True True False False --flip_test --shift_window_test --shift_size 16 --do_evaluate --ckpt_dir ckpt/kitti_swin_base.ckpt
  

• Evaluate with model (KITTI Swin-Large)

  $ python3 test.py --dataset kitti --kitti_crop garg_crop --data_path ../data/ --max_depth 80.0 --max_depth_eval 80.0 --backbone swin_large_v2 --depths 2 2 18 2 --num_filters 32 32 32 --deconv_kernels 2 2 2 --window_size 22 22 22 11 --pretrain_window_size 12 12 12 6 --use_shift True True False False --flip_test --shift_window_test --shift_size 16 --do_evaluate --ckpt_dir ckpt/kitti_swin_large.ckpt
  

Citation

@article{xie2023darkmim,
  title={Revealing the Dark Secrets of Masked Image Modeling},
  author={Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, Yue Cao},
  journal={arXiv preprint arXiv:2205.13543},
  year={2022}
}

Acknowledge

Our code is mainly based on GLPDepth[1]. The code of the model is from SwinTransformer[2] and Simple Baseline[3].

[1] Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth. [code]

[2] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. [code]

[3] Simple Baselines for Human Pose Estimation and Tracking. [code]