| Model | [email protected] (Original [1]) | [email protected] (Original [1]) | [email protected] (weight conversion) | [email protected] (weight conversion) | [email protected] (train) | [email protected] (train) |
|---|---|---|---|---|---|---|
| FCIS ResNet101 | 65.7 | 52.1 | 64.2 | 51.2 | 64.1 (1 GPU) | 51.2 (1 GPU) |
| Model | mAP/iou@[0.5:0.95] (original [1]) | mAP/[email protected] (original [1]) | mAP/iou@[0.5:0.95] (weight conversion) | mAP/[email protected] (weight conversion) | mAP/iou@[0.5:0.95] (train) | mAP/[email protected] (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.
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>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.
You can train the model with the following code.
Note that this code requires SciPy module.
python train_sbd.py [--gpu <gpu>]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>*0.0005You can download weights that were trained by ChainerCV.
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>*0.0005You can download weights that were trained by ChainerCV.
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.
- Yi Li et al. "Fully Convolutional Instance-aware Semantic Segmentation" CVPR 2017.