SAHI with line (LETR)

[aquexel]

Hello,

I'm using LETR for line detection with large pictures. Is it possible to use SAHI to split dataset (pictures too) with lines (labeled on Labelme) and do inference on large pictures ?

Thanks for all !

[fcakyon]

What is your label format? SAHI currently supports coco object detection and instance segmentation annotations.

[aquexel]

Hi
Actually format label is json like this:
{
"version": "4.0.0",
"flags": {},
"shapes": [
{
"label": "bus",
"points": [
[
260.936170212766,
23.33306055646483
],
[
193.936170212766,
20.33306055646483
],
[
124.93617021276599,
40.33306055646483
],
[
89.93617021276599,
102.33306055646483
],
[
81.93617021276599,
151.33306055646483
],
[
108.93617021276599,
146.33306055646483
],
[
88.93617021276599,
245.33306055646483
],
[
89.93617021276599,
323.33306055646483
],
[
116.93617021276599,
368.33306055646483
],
[
158.936170212766,
369.33306055646483
],
[
165.936170212766,
338.33306055646483
],
[
347.936170212766,
336.33306055646483
],
[
349.936170212766,
370.33306055646483
],
[
391.936170212766,
374.33306055646483
],
[
403.936170212766,
336.33306055646483
],
[
425.936170212766,
333.33306055646483
],
[
421.936170212766,
282.33306055646483
],
[
428.936170212766,
253.33306055646483
],
[
428.936170212766,
237.33306055646483
],
[
409.936170212766,
221.33306055646483
],
[
409.936170212766,
151.33306055646483
],
[
430.936170212766,
144.33306055646483
],
[
433.936170212766,
113.33306055646483
],
[
431.936170212766,
97.33306055646483
],
[
408.936170212766,
91.33306055646483
],
[
395.936170212766,
51.33306055646483
],
[
338.936170212766,
26.33306055646483
]
],
"group_id": null,
"shape_type": "lines",
"flags": {}
}
],
"imagePath": "2011_000025.jpg",
"imageData": null,
"imageHeight": 375,
"imageWidth": 500
}

[fcakyon]

That format is not supported, you can use the utilities from sahi to implement your custom scripts.

[aquexel]

Okay thank you,
If I convert lines to polygon with two points in COCO format will it work for inference and split ?
I don't know how the box will it work with lines
thanks

[fcakyon]

Check this url for object detection format: https://cocodataset.org/#format-data

You can represent lines as polygons (at least 3 points required), and set the segmentation field with point coords.

[aquexel]

Thanks,
I check that