https://github.com/836304831/keras-faster-rcnn-1

keras实现faster rcnn，end2end训练、预测; 持续更新中,见todo... ;欢迎试用、关注并反馈问题

https://github.com/836304831/keras-faster-rcnn-1

Repository

keras实现faster rcnn，end2end训练、预测; 持续更新中,见todo... ;欢迎试用、关注并反馈问题

https://github.com/836304831/keras-faster-rcnn-1/blob/master/

# keras-faster-rcnn

keras  [Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks](https://arxiv.org/pdf/1504.08083.pdf) ;[Mask RCNN](https://github.com/matterport/Mask_RCNN); Pascal VOC;

1. [](#)
2. [](#)
3. [](#)
4. [](#)
5. [](#)
6. [](#)
7. [toDoList](#toDoList)
8. [](#)

## :

a.resnet50;Conv1~4;Conv5rcnn; Conv3~Conv5;Batch Normalizationbatch size

b.720\*720; 720,,padding;600;1000

c.Keras Layer anchorsrpntargetroi2proposalsdetect boxes

d.10gtrpngtanchor

e.anchoryolo

f. rpn**256**anchors1:1PASCAL VOCGT**3**rpnanchors**21**rpn**80**anchors

g. rpn2000proposalsrcnnepoch300proposals,proposalsiou**0.9****0.7**




## 

Keras 2.2.4

tensorflow-gpu 1.9.0



## 

a) 

```shell
git clone https://github.com/yizt/keras-faster-rcnn.git
```

b) pascal voc,

```shell
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar
```

c) resnet 50

```shell
wget https://github.com/fchollet/deep-learning-models/releases/download/v0.2/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5
```

d) faster_rcnn/config.py

```python
pretrained_weights = '/opt/pretrained_model/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5'
voc_path = '/opt/dataset/VOCdevkit'
```

e) anchors
```shell
python gt_cluster.py --clusters 9
```
   ():
```shell
h:[53.25, 83.97, 149.73, 187.92, 202.07, 270.18, 357.79, 445.08, 488.2] 
w:[40.64, 128.43, 67.18, 304.62, 141.64, 539.33, 188.27, 336.6, 588.34]
ratio:[1.31, 0.65, 2.23, 0.62, 1.43, 0.5, 1.9, 1.32, 0.83]
ious:{1: 0.68557111415695, 2: 0.7254386539924431, 3: 0.692596649433563, 4: 0.6834624200806161, 5: 0.6157859655510957, 6: 0.6595932146657101, 7: 0.6685312631536455, 8: 0.665801489785847, 9: 0.6669742441576416, 10: 0.6121460291384846, 11: 0.6710420495151501, 12: 0.7141121719703986, 13: 0.720584796915154, 14: 0.5639737876945834, 15: 0.7321887018308443, 16: 0.6533920906006669, 17: 0.6594010218140989, 18: 0.5846891288412122, 19: 0.7055301287233066, 20: 0.6066710334483025}
mean iou:0.6643742977734857
```
   config.pyanchors
```python
    RPN_ANCHOR_HEIGHTS = [52.42, 85.64, 143.89, 186.92, 208.26, 266.1, 359.72, 446.26, 484.92]
    RPN_ANCHOR_WIDTHS = [40.85, 132.84, 66.24, 294.26, 135.53, 533.3, 190.26, 339.55, 591.88]
```
f) end2end
```shell
python train.py
```

## 

a) 

    PASCAL VOC 2007 [frcnn-resnet50.035.h5](https://drive.google.com/file/d/1TBGDTpdvCwGhIVEbv4t9q2mpzhL-5HdM/view?usp=sharing)


b) 

```shell
python inference.py --weight_path /tmp/frcnn-resnet50.035.h5
```

;demo_images

![examples](demo_images/inferece_examples.2.png)


## 



```shell
python evaluate.py --data_set test --weight_path /tmp/frcnn-resnet50.035.h5
```



```shell
 4952 ,1048.401778936386 
ap:{0: 0.0, 1: 0.6768214760548549, 2: 0.6760312377088041, 3: 0.6122568296243334, 4: 0.6367414618873344, 5: 0.6769495102412151, 6: 0.6631589701474099, 7: 0.5463179678366642, 8: 0.3078317216556386, 9: 0.4427729076167924, 10: 0.48129418164837967, 11: 0.5215081042808085, 12: 0.7005935007417052, 13: 0.5352984925242024, 14: 0.3158921077398846, 15: 0.7757831017666008, 16: 0.6199472930246124, 17: 0.32039103004250324, 18: 0.3694302780767011, 19: 0.659186483792924, 20: 0.6523787880767623}
mAP:0.5595292722244065
1052.8664383888245 
```

## 


1. gt_numGT 
2. positive_anchor_numrpnanchor
3. negative_anchor_numrpnanchor
4. rpn_miss_gt_numrpnanchorGT
5. rpn_gt_min_max_ioumax_iouGTanchorIoU;rpn_gt_min_max_ioumini-batchmax_iou;anchor
6. roi_numproposalnmsrcnnproposal
7. positive_roi_numrcnn
8. rcnn_miss_gt_numrcnn0.5iouGTproposals
9. rcnn_miss_gt_num_as(0.25), GTproposals
10. gt_min_max_iou: max_iouGTproposalsIoU;gt_min_max_ioumini-batchmax_iou;rpn                  


## toDoList
0. ()
1. Anchors()
2. batch_slice()
3. ,()
4. GT boxes()
5. 
6. indicestf.int64;float32()
7. GT boxesanchors()


## 
1. VOC 2007mAP55.9%(78%+)BN(batch size)? 

Owner

GitHub Events