Train_on_AF01.md

DesaySV-DL

Train on AF01 dataset on SSD^[1]

Prepare Training Data

1. Download AF01 image data

Put xml files into ~/data/AF01/"dataset"/Annotations and put image files into $HOME/data/AF01/"dataset"/BMPImages respectively. /dataset is the folder name of the AF01 dataset, e.g. $HOME/data/AF01/20170112/Annotations and ~/data/AF01/20170112/BMPImages.

2. Create image lists

This step is to create three text files: trainval.txt, test.txt, and test_name_size.txt in $CAFFE_ROOT/data/AF01/20170112. The dataset has been separated into trainval and test.

$ cd $CAFFE_ROOT/data/ && mkdir AF01
$ cp -r VOC0712/* AF01/

Contact me for the files create_list.sh and create_data.sh and replace them in $CAFFE_ROOT/data/AF01.

cd $CAFFE_ROOT/data/AF01
./create_list.sh

3. Convert images to lmdb

Create lmdb files for trainval and test with encoded original image and make soft links at $CAFFE_ROOT/examples/AF01_20170112/. The generated database files will be in $HOME/data/AF01/20170112/lmdb/AF01_20170112_trainval_lmdb and $HOME/data/AF01/20170112/lmdb/AF01_20170112_test_lmdb.

./create_data.sh

4. Create mean value

Image pre-processing like data normalization is one of the critical factors that affect the training process. To do data normalization, one way is to use the mean image.

Train

1. Download Network

Download the pre-trained SSD network weights from SSD github, unzip and copy the file VGG_VOC0712_SSD_300x300_iter_120000.caffemodel to $CAFFE_ROOT/models/VGGNet.

2. Train

Modify training script ssd_train_af01.py and train:

$ cd $CAFFE_ROOT
$ python examples/ssd/AF01/ssd_train_af01.py

3. Test

If you only want to test your trained network, run:

$ python examples/ssd/AF01/ssd_test_af01.py

^[1] SSD(SSD: Single Shot MultiBox Detector)