README.md

Faster RCNN Pytorch

re-implementation of faster rcnn (NIPS2015)

Please refer to https://arxiv.org/abs/1506.01497

Dataset

• VOC
• COCO

Data Augmentation (for implementation of original paper)

• Resize
• Horizontal Flip

Training Setting

• Use approximate joint training

- batch size : 1
- optimizer : SGD
- epoch : 13 
- initial learning rate 0.001
- weight decay : 5e-4
- momentum : 0.9
- scheduler : cosineannealing LR (min : 5e-5)

Results

• VOC

methods	Traning	Testing	Resolution	AP50
papers	2007	2007	**	69.9
papers	2007 + 2012	2007	**	73.2
this repo	2007	2007	**	71.00(+1.10)
this repo	2007 + 2012	2007	**	75.81(+2.61)

• COCO

methods	Traning	Testing	Resolution	mAP@[.5 .95]
papers	train	COCOval	**	21.2
papers	trainval	COCOval	**	-
this repo	COCOtrain2017	minival	**	20.7(-0.5%)
exp1 - more train	COCOtrain2017	minival	**	24.1(+2.9%)

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.207
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.387
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.199
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.039
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.237
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.352
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.209
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.294
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.298
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.055
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.354
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.510

exp1 - more train with new lr and wd

like torchvision environments,

- batch size : 4 (using 4 gpu)
- optimizer : SGD
- epoch : 26 
- initial learning rate 0.002
- weight decay : 1e-4
- momentum : 0.9
- scheduler : MultiStepLR - [16, 22]

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.241
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.463
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.228
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.091
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.288
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.376
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.225
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.326
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.331
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.122
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.398
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.503

** A way to resize frcnn is to make the image different size if the original image is different.

Quick Start for test

1 - download pth.tar files

• VOC
• faster_rcnn_voc.best.pth.tar here (about 1GB)
• COCO
• faster_rcnn_coco.best.pth.tar here (about 1GB)

2 - put tar file in like this (in saves)

dataset
evaluation
figures
logs
    |-- faster_rcnn_voc
        |-- saves
            |-- faster_rcnn_voc.best.pth.tar    
    |-- faster_rcnn_coco
        |-- saves
            |-- faster_rcnn_coco.best.pth.tar
anchor.py
...
main.py
...
utils.py

3 - set root and run test.py

test.py --config ./config_files/faster_rcnn_voc_test.txt
test.py --config ./config_files/faster_rcnn_coco_test.txt

Quick Start for demo

1 - run demo.py for demo : demo at demo figures voc or coco

demo.py --config ./config_files/faster_rcnn_coco_demo.txt
demo.py --config ./config_files/faster_rcnn_voc_demo.txt

Quick Start for train

1 - set your data root at config

config files is as follows

# name
name = faster_rcnn_coco

# data 
data_root = put your root in this part
data_type = coco

# training
epoch = 13
batch_size = 1

# testing
thres = 0.05

2 - run main.py for train

main.py --config ./config_files/faster_rcnn_coco_train.txt
main.py --config ./config_files/faster_rcnn_voc_train.txt

Process of faster rcnn

Citation

If you found this implementation and pretrained model helpful, please consider citation

@misc{csm-kr_Faster_RCNN_Pytorch,
  author={Sungmin, Cho},
  publisher = {GitHub},
  title={Faster_RCNN_Pytorch},
  url={https://github.com/csm-kr/Faster_RCNN_Pytorch//},
  year={2022},
}

at resize transform, create target['size'] (h, w)

todo

• mkae synchro test like https://github.com/csm-kr/detr_pytorch/blob/master/test.py