Given a picture with a bird, we are supposed to box the bird.
In src/data directory, images.txt is the index of all images, bouding_boxex.txt is the label box of all images and images contains all images. Box data make up of 4 data: the top left corner coordinate of box, width of box and height of box.
For traditional CNN and FC, it will meet degeneration problems when layers go deep.
In paper Deep Residual Learning for Image Recognition, they try to solve this problem by using a Residual Block:
These blocks compose ResNet:
I use ResNet-18 in this project by adding a 4-dimension layer after ResNet-18 to predict box's x, y ,w and h.
Loss: smooth l1 loss
Metric: IoU of groound truth and prediction, threshold=0.75
Resize all images to 224*224*3
Then normalize and standardize all pixel channel.
Split all data into 9000 training data and 2788 tesing data. Train network on training data using batch size=128, epoch=100 and validation split ratio=0.1
Training result:
Testing result:
Red box represents ground truth and green box is the prediction of network.
Failed example:
You should keep the directory structure.
python 3.6
tensorflow 1.3.0
keras 2.1.0
numpy
PIL
pickle
matplotlib
In src directory:
python getdata.py to preprocess data.
If you want to train model, python train.py
If you want to test on trained model(if you had trained model), python test.py
Deep Residual Learning for Image Recognition: https://arxiv.org/pdf/1512.03385.pdf
CKCZZJ
MIT