alankbi
Any constraints on Input Image Size for Custom Object Training
I am able to train on the given sample dataset of dog. But not on my data is a balance sheet and income statement (will be having classes as table and cell) which has the size of (1125, 2000, 3), not able to share the image here because of confidentiality. Is there any restriction for image shape which model does not allow us to train?
When I run the training I do not get any error. But output is empty array and when checked for loss it is nan
I have tried with both way i.e. with csv file and without csv file where I was able to get output for dog but not for my document image
Could you give us more information? From what I understand, your images have 3 channels, width 1125 and height 2000. Do you have an XML file (for each image) or a CSV file containing all other information such as boxes, classes etc.?
Maybe you could clarify what you are trying to detect in those images? Is it image classification (detect whether it is a balance sheet or income statement) or are you trying to detect specific objects (in that case some paragraphs, headlines etc.)?
I came to know that the error was due to images generated with grayscale=True flag by pdf2image python module. I am not using image classification. I want to detect objects from the document which are cell and table.
I have two classes 1. table (for detecting table bounding box in document image) 2. cell (to detect cell bounding box in the detected table)
Would like to know is detecto capable of doing this? i.e. to train the table and cell object detector ? 1 table can have 180 cells. If i am getting correct bounding boxes, I can figure out how to get the structure of the table but need correct bboxes around the cell and table