Please can you explain what are those parameters important for

[ghost]

    "min_input_size":      
    "max_input_size":      
    "anchors":            

[darjoo]

min_input_size defines the minimum size of image during training. max_input_size defines the maximum size of image during training anchors These anchors are generated from your dataset with k-means clustering, they are used for transforming the predictions to actual bounding boxes.

Do note that if you choose input sizes that vary largely from your actual image size, you may lose important data/gain unnecessary or weird information due to image scaling.

[ghost]

@darjoo Can you confirm if its right for the input size : input_size = total image width total image height ?? or boundbox width boundbox height ??

for anchors: how can I get those numbers from the dataset ???

[darjoo]

No, input_size is not width * height. Input size is a single number representing both width and height individually. It is a single number due to the assumption that we're dealing with a square image. In general, most object detection problems will have no issues with rectangular images being resized into a square.

Use gen_anchors.py to generate the anchors from your dataset.