Crop image but still mapping annotation

[winter2897]

Hi @ashnair1 ,
I want to crop the large image into several small images and retain the annotation mapping. Can you give me advice?

[ashnair1]

I'm not sure what you mean by annotation mapping. Is the annotation made at the large image level? So for example, you have one large image and one annotation file and you want to know how to tile it in order to make it a COCO style dataset?

[winter2897]

I already have the label file of the large image. I want it when I cut it into small pictures. These small pictures are automatically labeled in the COCO format. @ashnair1

[ashnair1]

What format is the label file?

It seems you're working on geospatial imagery so I'm assuming your image is a GeoTIFF and label file is a GeoJSON/Shapefile. Is that correct?

[winter2897]

I'm working on large image sizes of solar panels. I label them with labelme and then use the code labelme2coco.py

to convert the .json files to COCO format which is a single file train.json.

[winter2897]

Since the size image is too large, I want to cut it into small sheets. These small photos, after being cut, must automatically be labeled in the COCO format.

[ashnair1]

@winter2897
Unfortunately I don't know of any tools that will do what you want directly.

The most obvious solution would be:

1. Crop images to specific size (for example 256 x 256)
2. Label the image tiles using labelme and convert them to COCO format using label2coco.py

This means that you will need to do the annotation again but depending on your situation it might be easier.

Alternatively, if you don't want to do the annotation, you will need to write a script to modify the annotation

One way you could do this:

1. Tile images to specific size (for example 256 x 256)
2. Convert your large COCO annotation into a segmentation mask and tile that to size 256 x 256
3. Now try converting segmentation mask to coco annotation

[ashnair1]

You could refer det2seg.py for getting an idea on how to convert COCO dataset to segmentation masks.

[winter2897]

Hi @ashnair1 ,
I just got back to work. Thank you for giving me advice. I am making the way you instruct. However it is experiencing this problem.

[winter2897]

My class: [{'supercategory': 'Dust1', 'id': 0, 'name': 'Dust1'}, {'supercategory': 'Dust2', 'id': 1, 'name': 'Dust2'}, {'supercategory': 'Dust3', 'id': 2, 'name': 'Dust3'}, {'supercategory': 'Wire', 'id': 3, 'name': 'Wire'}]. I don't have class 'supercategory': 'Background'

[ashnair1]

You don't need to have Background as a class.

Edit: Your ids should actually start from 1.

[winter2897]

Hi @ashnair1,

if colour_map.shape != (len(cats) + 1, 3):
        raise AssertionError("Incorrect shape of color map array")

Why len(cats) + 1.....? Can you explain it to me?

[ashnair1]

The +1 is for background. The palette will have len(cats) + 1 unique colors for categories and background (which is black by default)

[winter2897]

I removed parameter 1 from (len(cats) + 1, 3), the result showed that the label Dust1 was the same color as Backgound. The other labels are still correctly illustrated.

[winter2897]

opp, okay. I will fix the id in the .json file.

[winter2897]

Dust1 and Wire still missing

[winter2897]

"categories": [
{
"supercategory": "Dust1",
"id": 1,
"name": "Dust1"
},
{
"supercategory": "Dust2",
"id": 2,
"name": "Dust2"
},
{
"supercategory": "Dust3",
"id": 3,
"name": "Dust3"
},
{
"supercategory": "Wire",
"id": 4,
"name": "Wire"
}

[winter2897]

I fixed it. Thank you very much @ashnair1 . I have to edit "category_id" accordingly.

[winter2897]

[fcakyon]

@winter2897 you can use https://github.com/obss/sahi library for cropping coco annotations and images.

• Slice a coco formatted dataset:

from sahi.slicing import slice_coco

coco_dict, coco_path = slice_coco(
    coco_annotation_file_path=coco_annotation_file_path,
    image_dir=image_dir,
    slice_height=256,
    slice_width=256,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2,
)