FPN

Pytorch Implementation of "Feature Pyramid Networks for Object Detection"
You can star this repository to keep track of the project if it's helpful for you, thank you for your support.

Environment

OS: Ubuntu 16.04
Python: python3.x with torch==1.2.0, torchvision==0.4.0

Performance

Backbone	Train	Test	Pretrained Model	Epochs	Learning Rate	RoI per image	AP
Res50-FPN	trainval35k	minival5k	Pytorch	12	2e-2/2e-3/2e-4	512	35.5
Res101-FPN	trainval35k	minival5k	Pytorch	12	2e-2/2e-3/2e-4	512	37.4

Trained models

You could get the trained models reported above at 
https://drive.google.com/open?id=1xm8z-EMbNG17sQzd-2FRRLVk_N7UIOhE

Usage

Setup

cd libs
sh make.sh

Train

usage: train.py [-h] --datasetname DATASETNAME --backbonename BACKBONENAME
                [--checkpointspath CHECKPOINTSPATH]
optional arguments:
  -h, --help            show this help message and exit
  --datasetname DATASETNAME
                        dataset for training.
  --backbonename BACKBONENAME
                        backbone network for training.
  --checkpointspath CHECKPOINTSPATH
                        checkpoints you want to use.
cmd example:
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train.py --datasetname coco --backbonename resnet50

Test

usage: test.py [-h] --datasetname DATASETNAME [--annfilepath ANNFILEPATH]
               [--datasettype DATASETTYPE] --backbonename BACKBONENAME
               --checkpointspath CHECKPOINTSPATH [--nmsthresh NMSTHRESH]
optional arguments:
  -h, --help            show this help message and exit
  --datasetname DATASETNAME
                        dataset for testing.
  --annfilepath ANNFILEPATH
                        used to specify annfilepath.
  --datasettype DATASETTYPE
                        used to specify datasettype.
  --backbonename BACKBONENAME
                        backbone network for testing.
  --checkpointspath CHECKPOINTSPATH
                        checkpoints you want to use.
  --nmsthresh NMSTHRESH
                        thresh used in nms.
cmd example:
CUDA_VISIBLE_DEVICES=0 python test.py --checkpointspath fpn_res50_trainbackup_coco/epoch_12.pth --datasetname coco --backbonename resnet50

Demo

usage: demo.py [-h] --imagepath IMAGEPATH --backbonename BACKBONENAME
               --datasetname DATASETNAME --checkpointspath CHECKPOINTSPATH
               [--nmsthresh NMSTHRESH] [--confthresh CONFTHRESH]
optional arguments:
  -h, --help            show this help message and exit
  --imagepath IMAGEPATH
                        image you want to detect.
  --backbonename BACKBONENAME
                        backbone network for demo.
  --datasetname DATASETNAME
                        dataset used to train.
  --checkpointspath CHECKPOINTSPATH
                        checkpoints you want to use.
  --nmsthresh NMSTHRESH
                        thresh used in nms.
  --confthresh CONFTHRESH
                        thresh used in showing bounding box.
cmd example:
CUDA_VISIBLE_DEVICES=0 python demo.py --checkpointspath fpn_res50_trainbackup_coco/epoch_12.pth --datasetname coco --backbonename resnet50 --imagepath 000001.jpg

Reference

[1]. https://github.com/jwyang/fpn.pytorch
[2]. https://github.com/open-mmlab/mmdetection