Tensorflow1.14 object detection anchor free
anchor point with center sample, every box with different colors.
Train on 1 gtx1660 with 6G memory, in PASCAL VOC07 trainval and test dataset.I found centerness has less impact when imports center sample.You can download the trained weight in Baidu driver link, password: f4e5 and get the centerness weight, put them in checkpoint files and run evay.py. The without centerness weight in Baidu driver link, password: ksm8. You can set centerness in config.py.
| Method | mAP(0.5) |
|---|---|
| centerness(512x640px) | 67.6 |
| without centerness(512x640px) | 67.4 |
Download resnet50's weight from tensorflow's official model zoo. ResNet-50
Firstly, convert your lables into xml. Get the label names and write them into one txt file. For example, I get labels 1.xml, 2.xml, .... and the txt file shoud be
1
2
3
Then config the config.py
Finally, Run train.py, Run <tensorboard --logdir=F:\back_up\FCOS\checkpoint --host=127.0.0.1> to get tensorboard, then you can check loss curves and detection results.
Config the config.py about the val_restore_path Then config and run demo.py
Config the config.py about the val_restore_path Then run eval.py. Notice that batch size should be setted as 1
I mainly refer @zhenghao977 and convert (His/Her) some codes into tensorflow from pytorch.
- Improve Data Prepare Pipeline
- COCO Results
- fp16 training
- Export .pb Model
- Export Trt Engine
- And So On