With this project I tried to segment AR markers, even if they are not detectable. The network is based on SegNet. The publication is SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation.
After a training on synthetical data, the network was able to provide an acceptable result on unseen synthetic data with a IoU of 79,56.
But the result on real images is not very impressive, as the following examples shows.
Even if the project does not meet the expectations, the implementation can be used for further projects.
There are two ways to train the net.
On the one hand with synthetic data. To do this, each second iteration is switched between the following two samples. Where one is generated completely synthetically and the other with a random background image from ./bg.
For the above example the training was started as follows:
python3 train_synthetic_data.py -i 150000 -bs 4 -APRILTAG_36h11
A list with all arguments can be seen with the -h parameter.
I have already tried to train SegNet on real data, but the generation of a enough large dataset was too time-consuming for me, so I don't have any results yet.
python3 train_real_data.py -i 150000 -bs 4 -dp ./train_dataset
A list with all arguments can be seen with the -h parameter.
If you want to try this, you can use the train_real_data.py script.
However, a calibration for the camera is required, as done here.
You can run a prection with the following command:
python3 predict.py -i ./train_dataset/0.jpg -o ./0_prediction.jpg
A list with all arguments can be seen with the -h parameter.