What about? Instance segmentation model : Mask-RCNN !

So what : Improved performance!

input

Mask-RCNN Caculates....

output

Mask rcnn transfrom layer Interpolation experiments

Before Image goes into mask rcnn backbone module , image is resized to fixed size to get good features. When resizing image, generally Bilinear interpolation is used. There're many kind of interpolation methods to resize in this repository code.

Train

main options

• --dt : dataset

  • pf : pennFudan
  • bln : balloon

• --model : interpolation method

  • bicubic (👍)
  • bilinear
  • nearest

Usage

python train.py --dt pf --model  bicubic -o ./model/something.pth

Secondary options

• --out : default = './model/new.pth'

• --epochs : default = 50

• --batch : default = 4

• --device : default = 'cuda:0'

• --workers : default = 4

Evaluate

About performance

• mAP
  • mask
  • bbox

Jupyter (interactive)

: evaluate.ipynb

Shell

python evaluate.py -m /path/to/modelA.pth /path/to/modelB.pth -o /where/to/save/figure_dir

options

• -m ,--model (default) ./models/*.pth
• -o ,--output (default) false (false : prints evaluation results on console, true : saves graph images in ./results directory )

e.g )

python evaluate.py 
python evaluate.py -m /path/to/modelA.pth /path/to/modelB.pth -o true
python evaluate.py -o true
python evaluate.py -m /path/to/modelA.pth /path/to/modelB.pth 

image output : modelName_ap_epochs.jpg , modelName_ap_table.jpg

Inference

Adjust bicubic mask rcnn to your image.

Bash Usage

python inference.py -m ./models/pf_4_nearest.pth -i ./input.jpg -o ./output.jpg

options

• -m ,--model (default) './models/pf_4_bicubic.pth'
• -i ,--input (default) './sample/pds1.jpg'
• -o ,--output (default) './sample/pre_pds1.jpg'

ETC

If you have a question , feel free to ask me.