fpn

Examples of Faster R-CNN and related models using FPN [1, 2]

Performance

MS COCO 2017 Val

Backbone	Supervision	Original (Bbox)	Ours (Bbox)	Original (Mask)	Ours (Mask)
FPN w/ ResNet50	bbox	36.7 % [3]	37.1 %
FPN w/ ResNet101	bbox	39.4 % [3]	39.5 %
FPN w/ ResNet50	bbox + mask	37.3 % [3]	38.0 %	33.7% [3]	34.2 %
FPN w/ ResNet101	bbox + mask	39.4 % [3]	40.4%	35.6% [3]	36.0%

Scores are the mean of mean Average Precision (mmAP).

Demo

If faster_rcnn_* is used as --model, the script conducts object detection. If mask_rcnn_* is used as --model, the script conducts instance segmentation instead. This demo downloads MS COCO pretrained model automatically if a pretrained model path is not given.

$ python demo.py [--model faster_rcnn_fpn_resnet50|faster_rcnn_fpn_101|mask_rcnn_fpn_50|mask_rcnn_fpn_101] [--gpu <gpu>] [--pretrained-model <model_path>] <image>.jpg

Evaluation

For object detection, use chainercv/examples/detection/eval_detection.py for evaluation. For instance segmentation, use chainercv/examples/detection/eval_instance_segmentation.py for evaluation.

Train

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

$ mpiexec -n <#gpu> python train_multi.py [--model faster_rcnn_fpn_resnet50|faster_rcnn_fpn_resnet101|mask_rcnn_fpn_resnet50|mask_rcnn_fpn_resnet101] [--batchsize <batchsize>]

Note that cv2 is required for training Mask R-CNN.

References

1. Tsung-Yi Lin et al. "Feature Pyramid Networks for Object Detection" CVPR 2017
2. Kaiming He et al. "Mask R-CNN" ICCV 2017
3. Detectron