We are often training multiple models that search for specific things in an image. When found, they are passed to other models of their corresponding category.
For example :
First model's job is to find the "big" objects and then pass it to the right second model specificly trainned for this big image class to find the "smaller-parts". This is a 2 level nesting, but we should be able to add more!
I didn't find an annotation format / tool that supports that yet. Also, a change to "parent" object should be changing (or not) the children's position following the user defined strategy. For example, resizing "left" side of a bounding box for -15px should translate all the children positions for -15px if the nested change strategy is telling to do so.
Here is a basic example of nested annotations :
{
"filename": "12345.jpg",
"key": "12345",
"size": {
"width": 4608,
"height": 2240,
"depth": 3
},
"annotations": {
"ground_truth": {
"objects": [
{
"name": "building",
"bndbox": {
"left": 600,
"top": 100,
"right": 2000,
"bottom": 1000
},
"size": {
"width": 1400,
"height": 900,
"depth": 3
},
"annotations": {
"ground_truth": {
"objects": [
{
"name": "window",
"bndbox": {
"left": 100,
"top": 34,
"right": 853,
"bottom": 130
}
},
{
"name": "window",
"bndbox": {
"left": 100,
"top": 34,
"right": 853,
"bottom": 130
}
},
{
"name": "door",
"bndbox": {
"left": 100,
"top": 134,
"right": 430,
"bottom": 230
}
},
{
"name": "door",
"bndbox": {
"left": 100,
"top": 134,
"right": 430,
"bottom": 230
}
}
]
}
}
}
]
}
}
}
We are often training multiple models that search for specific things in an image. When found, they are passed to other models of their corresponding category.
For example :
First model's job is to find the "big" objects and then pass it to the right second model specificly trainned for this big image class to find the "smaller-parts". This is a 2 level nesting, but we should be able to add more!
I didn't find an annotation format / tool that supports that yet. Also, a change to "parent" object should be changing (or not) the children's position following the user defined strategy. For example, resizing "left" side of a bounding box for -15px should translate all the children positions for -15px if the nested change strategy is telling to do so.
Here is a basic example of nested annotations :