Skip to content

Latest commit

 

History

History
 
 

fcis

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 

Example codes for FCIS [1]

Performance

SBD Train & Test

Model mAP@0.5 (Original [1]) mAP@0.7 (Original [1]) mAP@0.5 (weight conversion) mAP@0.7 (weight conversion) mAP@0.5 (train) mAP@0.7 (train)
FCIS ResNet101 65.7 52.1 64.2 51.2 64.1 (1 GPU) 51.2 (1 GPU)

COCO Train & Test

Model mAP/iou@[0.5:0.95] (original [1]) mAP/iou@0.5 (original [1]) mAP/iou@[0.5:0.95] (weight conversion) mAP/iou@0.5 (weight conversion) mAP/iou@[0.5:0.95] (train) mAP/iou@0.5 (train)
FCIS ResNet101 29.2 49.5 27.9 46.3 24.3 (3 GPU) 42.6 (3 GPU)

*: We use random sampling for sampling strategy. The original paper used OHEM sampling strategy.

Demo

Segment objects in an given image. This demo downloads SBD pretrained model automatically if a pretrained model path is not given.

python demo.py [--dataset sbd|coco] [--gpu <gpu>] [--pretrained-model <model_path>] <image.jpg>

Evaluation

The evaluation for sbd dataset can be conducted using chainercv/examples/instance_segmentation/eval_sbd.py and the one for coco dataset can be conducted using chainercv/examples/instance_segmentation/eval_coco.py.

Train

You can train the model with the following code. Note that this code requires SciPy module.

SBD Train with single GPU

python train_sbd.py [--gpu <gpu>]

SBD Train with multiple GPUs

If you want to use multiple GPUs, use train_sbd_multi.py. Note that this code requires chainermn module.

mpiexec -n <n_gpu> python train_sbd_multi.py --lr  <n_gpu>*0.0005

You can download weights that were trained by ChainerCV.

COCO Train with multiple GPUs

If you want to use multiple GPUs, use train_coco_multi.py. Note that this code requires chainermn module.

mpiexec -n <n_gpu> python train_coco_multi.py --lr  <n_gpu>*0.0005

You can download weights that were trained by ChainerCV.

Convert Mxnet model

Convert *.params to *.npz. Note that the number of classes and network structure is specified by --dataset.

python mxnet2npz.py [--dataset sbd|coco] [--out <npz filename>] <param filename>

You can download weights that were converted the script.

References

  1. Yi Li et al. "Fully Convolutional Instance-aware Semantic Segmentation" CVPR 2017.