I need to make an object detection model in which used my previously created labels in json format to feed into deep learning model. I checked the shared workflows on Knime website, but i would like to use not only class information but also shape information presented in label to feed into model.Is there a way to do that? I want to do object detection in my model, so i need to use region proposals.
Is there any Mask-R CNN workflow template on website to check and understand how to perform in Knime?
as far as I know there is not example workflow for a Mask-R CNN.
Are you planning to reuse a trained network or was your goal to train the network from scratch?
What use case are you trying to solve? Maybe an easier network structure like yolo might be sufficient and probably easier to set up?
I will start to train the network from scratch.I will create a model which identifies irregularities on a given image. Yolo also works for me. I just need an example workflow to start with.
But if i use yolo, i need to create labels in yolo format. Is there a way to create labels in yolo format in Knime ?
I created a couple of deep learning example use cases in the past and would be happy to support you, but I don’t have a yolo example yet that I could share with you.
Are you planning to use a public dataset to train your model? If yes I’m happy to have a look at the data and think about a possible workflow setup.
I will not use a public dataset, and it is not allowed to share data unfortunately. However, I have a dataset which has gray background and i want to detect some deformation(brighter local regions) on surface .Therefore, i need to use an object detection algorithm with json or yolo labeling.If you have an example that you can share with me, it would be to helpful for me to understand basic needs to build this kind of model on Knime.
Only example workflows i found out on Knime website mostly deal with image classification. In which they only feed the class category in label information into model. However, i need to use object detection algorithms with labels in which there are coordinate informations of deformed local regions.
I have also some question marks about preprocessing steps. I have a grey scale image but still do i need to normalize image or use some threshold methods to binarize the image and get more accurate results? At least, before labeling, maybe i need to use some image processing before using segmentation methods.
I recommend to normalize the data between, so the different shades of grey are encoded the with values between 0 and 1.
Attached is a workflow where I worked on a similar use case using a Unet architecture. In that case I had the input images and label images. Please have a look and let me know in case of any questions and whether this was helpful or not.
In addition you can find here an example where I trained a network with multiple outputs and a related blog post. In case you a have fixed number of output values describing your bounding box in addition to the class label this might be an other option.
Many thanks for your time,you’re a great help, I must say. I just try to understand the U-net architecture that you created. Could you please also give some info about how you create the labelings or preprocessing steps that you followed? Then,i guess you imported labels and images directly as table.
In my case the labels where provided as images already. So I didn’t do much preprocessing besides normalizing the images with the Image Calculater node.