This repository is not maintained. Please use ChainerCV.

Feature Pyramid Networks for Object Detection

This is an implementation of FPN (Feature Pyramid Networks) using Chainer

Performance

mmAP on COCO 2014 minival

backbone	original (Detectron)	ours (inference only)	ours (train & inference)
ResNet50	36.7 %	35.7 %	37.1 %
ResNet101	39.4 %	38.2 %	39.5 %

Requirements

• Python 3.6
• Chainer 4.0+
• CuPy 4.0+
• ChainerCV (we need to build from master branch)
• ChainerMN 1.3
• pycocotools

Demo

$ curl -LO https://github.com/Hakuyume/chainer-fpn/releases/download/assets/faster_rcnn_fpn_resnet50_coco.npz
$ python3 demo.py [--gpu <gpu>] --model resnet50 --pretrained-model faster_rcnn_fpn_resnet50_coco.npz <image>

Training

$ mpiexec -n <#gpu> python3 train_coco.py --model resnet50

Our experiments were conducted with 8 GPUs.

Evaluation

$ python3 eval_coco.py [--gpu <gpu>] --model resnet50 --pretrained-model faster_rcnn_fpn_resnet50_coco.npz

or

$ python3 eval_coco.py [--gpu <gpu>] --model resnet50 --snapshot result/snapshot_iter_90000

Convert weights from Detectron

1. Download weights from Detectron's model zoo.

$ curl -L https://s3-us-west-2.amazonaws.com/detectron/35857345/12_2017_baselines/e2e_faster_rcnn_R-50-FPN_1x.yaml.01_36_30.cUF7QR7I/output/train/coco_2014_train%3Acoco_2014_valminusminival/generalized_rcnn/model_final.pkl -o e2e_faster_rcnn_R-50-FPN_1x.pkl
$ curl -L https://s3-us-west-2.amazonaws.com/detectron/35857890/12_2017_baselines/e2e_faster_rcnn_R-101-FPN_1x.yaml.01_38_50.sNxI7sX7/output/train/coco_2014_train%3Acoco_2014_valminusminival/generalized_rcnn/model_final.pkl -o e2e_faster_rcnn_R-101-FPN_1x.pkl

2. Convert weights.

$ python3 detectron2npz.py e2e_faster_rcnn_R-50-FPN_1x.pkl faster_rcnn_fpn_resnet50_coco.npz

Note: Since the mean value in Detectron is different from that in ChainerCV, --mean=detectron option should be specified for converted weights.

$ python3 eval_coco.py [--gpu <gpu>] --model resnet50 --mean=detectron --pretrained-model faster_rcnn_fpn_resnet50_coco.npz

References

1. Tsung-Yi Lin et al. "Feature Pyramid Networks for Object Detection" CVPR 2017
2. Detectron
3. Mask R-CNN by @wkentaro (for the implementation of RoIAlign)