momuli / smenet

SMENet

This is a pytorch implementation of SMENet

Requirements

1. pytorch == 1.1.0

2. cuda 8.0

3. python == 3.7

4. opencv(CV2)

Data Prepare

1. Please download NWPU VHR-10
2. Convert to PASCAL VOC data format
3. Create dataset folder

./SMENet/VOCNWPU/

4.Data format

├── VOCNWPU
│   ├── VOC2012
│       ├── Annotations
│       ├── JPEGImages
│       ├── ImageSets
│         ├── train.txt
│         ├── test.txt
|         ├── val.txt
|         ├── trainval.txt

Demo

1.Please download weights file SMENet.pth, and put it to:

./SMENet/weights/

2. Run visual_SMENet.py:

cd ./SMENet/demo/visual_SMENet.py
modify parser.add_argument('--trained_model', default='../weights/SMENet.pth', type=str, help='Trained state_dict file path to open')
python visual_SMENet.py

Train

if you want to train your own dataset:

1. Convert your dataset to PASCAL VOC and put it to `./SMENet/dataset-file-name/`
2. Modify parameters  `HOME` and  `num_classes` in `./SMENet/data/config.py` :
    HOME= absolute path of the SMENet folder
3. Modify parameters `VOC_CLASSES` in `./SMENet/data/voc0712.py`
4. `python train_SMENet.py`
5. save `*.pth` to weights, like `./SMENet/weights/*.pth`

Eval

if you want to eval trained model:

1. cd ./SMENet/eval.py
2. Modify parser.add_argument('--trained_model', default='weights/*.pth', type=str, help='Trained state_dict file path to open')
3. python eval.py

smenet's Issues

how to download IPIU-M dataset, I don't find it

Data spliting ?

please can you tell us when train and test percentage did you use ?

Both oer and ose have been commented out

Both oer and ose have been commented out. Is there any version that has not been commented out

您如何将NWPU VHR-10数据集转成VOC格式的，方便给出转化形式的代码吗，非常感谢

负样本没有坐标和label信息，您是怎么产生的谢谢~

test

I hope to use the network model you designed to train the new dataset I created (3 categories). I did not modify the image size of the input network (using the initial 400Pixel in the code), but only modified the file used to create the dataset, voc2017.py. The final training loss was 0.899 (different learning rates were attempted multiple times), but the eval result was poor, with a mAP of only 0.0566. Excuse me, are there any areas where I haven't made the necessary modifications or omissions?
The visualization results of target detection are shown in the following figure. For the categories detected in the red box, they are correct, while others are error detection

I just rewrote the__ getitem__ ()

Issue about some module in your projects

The implementation of the OER module seems to be different from the one described in your paper,
In train_SMENet.py the OER module has been commented out,
In SMENet.py the OSE module has also been commented out,
why do you comment out these modules？Can this achieve better results？

Exception in other datasets during network training

I hope to use the network model you designed to train the new dataset I created (3 categories). I did not modify the image size of the input network (using the initial 400Pixel in the code), but only modified the file used to create the dataset, voc2017.py. The final training loss was 0.899 (different learning rates were attempted multiple times), but the eval result was poor, with a mAP of only 0.0566. Excuse me, are there any areas where I haven't made the necessary modifications or omissions?
The visualization results of target detection are shown in the following figure. For the categories detected in the red box, they are correct, while others are error detection

I just rewrote the__ getitem__ ()