Specific format of annotation #60

ycui123 · 2018-07-02T14:51:19Z

Could you please tell me the format of annotation? I generated my own dataset and I want to train it. I know exactly where my objects are in the image. In that case, I don't want to manually generate annotation. I can write some code to do it for me if I know the format of annotation for YOLO v3. Thank you

AlexeyAB · 2018-07-02T16:07:58Z

.txt-file for each .jpg-image-file - in the same directory and with the same name, but with .txt-extension, and put to file: object number and object coordinates on this image, for each object in new line: <object-class> <x> <y> <width> <height>

Where:

<object-class> - integer number of object from 0 to (classes-1)
<x> <y> <width> <height> - float values relative to width and height of image, it can be equal from (0.0 to 1.0]
for example: <x> = <absolute_x> / <image_width> or <height> = <absolute_height> / <image_height>
atention: <x> <y> - are center of rectangle (are not top-left corner)

For example for img1.jpg you will be created img1.txt containing:

1 0.716797 0.395833 0.216406 0.147222
0 0.687109 0.379167 0.255469 0.158333
1 0.420312 0.395833 0.140625 0.166667

https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

ycui123 · 2018-07-02T16:19:46Z

Thank you @AlexeyAB . This is very helpful.

ycui123 · 2018-07-11T13:44:10Z

Hi @AlexeyAB ,

I generated annotations by a script. But when I try to test my model. I cannot see a bounded box on my test data. Is it because I didn't use the tool to generate bounded box for my training data?

ycui123 · 2018-07-11T14:23:54Z

I only have one class and I trained for 4000 iterations. I used the command line to test.

darknet.exe detector test data/obj.data yolo-obj.cfg yolo-obj_4000.weights -thresh 0.25 E:\darknet\build\darknet\x64\data\obj_test\11.jpg -ext_output

I only got "E:\darknet\build\darknet\x64\data\obj\11.jpg: Predicted in 0.060051 seconds." without any bounded boxes and coordinates of objects.

Thank you .

ycui123 · 2018-07-11T15:21:40Z

Alright, I think the problem is the model is not predicting. That's weird. I trained for 4000 iterations and got 0.7 loss.

I'm not getting results even when I test on training data.

AlexeyAB · 2018-07-11T15:36:18Z

@ycui123

What mAP can you get? https://github.com/AlexeyAB/darknet#when-should-i-stop-training
Try to open your dataset in the Yolo_mark, and show screenshot.

ycui123 · 2018-07-11T16:34:22Z

Thank you for your reply. I think I found the bug. I labelled my object wrong somehow which leads to useless traning.

ycui123 · 2018-07-11T16:46:20Z

But does Yolo_mark resize the image before it does the mark?

Edited: My bad. It doesn't. I reversed x and y and it works! Thanks a lot

koutini · 2019-03-12T14:37:44Z

hello, i have annotation with three valuers (example 156 111 111) and i i don't understand how to convert annotation with three valuers to yolo format

sarratouil · 2019-03-12T15:24:58Z

@koutini i have the same problem

sarratouil · 2019-03-12T15:25:36Z

@koutini please help me to resolve it

koutini · 2019-03-12T15:28:03Z

@sarratouil thank you for the support

sarratouil · 2019-03-12T15:34:15Z

@AlexeyAB @ycui123 @ido-ran @RRMoelker how can i convert to format yolo this format of annotations
please help me

AlexeyAB · 2019-03-12T15:51:32Z

@koutini @sarratouil Hi,

What dataset do you use?
What do these 3 values mean? Where is: class_id, x, y, width and height?

sarratouil · 2019-03-12T16:06:12Z

@AlexeyAB i use this dataset to train to detect iris
i download this dataset from " https://web.inf.ufpr.br/vri/databases/iris-location-annotations/ "
it is annoted dataset but i hav'nt any idea about this 3 values

sarratouil · 2019-03-12T16:07:24Z

@AlexeyAB please help me to resolve it

koutini · 2019-03-12T16:10:08Z

@AlexeyAB thank you so much for your answer ,
we use iris data set annotation for that we can get always square to detect the position iris in an image , actually i don't understand what represents each number in this annotation

sarratouil · 2019-03-12T16:52:55Z

@AlexeyAB may be 1st ,2nd values are the coordinates of the centre of square
and 3rd is (width and height)

koutini · 2019-03-12T16:56:39Z

@sarratouil I told you

koutini · 2019-03-12T17:01:28Z

@AlexeyAB also may be 1st ,2nd values are coordinates of the first point when the writer click in the image and 3rd is width and height

AlexeyAB · 2019-03-12T18:04:13Z

@koutini @sarratouil

I just don't see the link to IRIS images.

So you should write your code on Bash/Python/C... for converting
from irisX irisY irisR
to 0 irisX/image_width irisY/image_height irisR/image_width irisR/image_height

Just for example, you can look at this script - how to read CSV files on bash: https://github.com/AlexeyAB/darknet/blob/master/scripts/windows/otb_get_labels.sh

sarratouil · 2019-03-13T10:24:53Z

@AlexeyAB
Thank you very much
I write a code with python and it work very good .

sarratouil · 2019-03-14T10:30:31Z

@AlexeyAB you have any idea how i can convert anymy yolo-tiny-obj that i trained it to TensorFlow to use on Android platforms

AlexeyAB · 2019-03-14T14:26:47Z

@sarratouil

Try to use this repository: https://github.com/thtrieu/darkflow there is about converting Yolo to TensorFlow: How to convert darknet model to tensorflow model thtrieu/darkflow#527
Or just use OpenCV for android https://opencv.org/releases.html and use your yolo-tiny-obj.cfg / yolo-tiny-obj.weights by using these examples:
- C++ https://github.com/opencv/opencv/blob/master/samples/dnn/object_detection.cpp#L192-L221
- Python https://github.com/opencv/opencv/blob/master/samples/dnn/object_detection.py#L129-L150

Also you can look at these repos:

koutini · 2019-03-17T21:28:40Z

dear @AlexeyAB hello ,
after the train with darknet and yolo the detection box it not exact my question is : anchors can be the raison of this problem ?

AlexeyAB · 2019-03-17T22:01:26Z

@koutini Hi,

What do you mean?
Can you show screenshot of example?

koutini · 2019-03-17T22:11:27Z

cindyweng · 2020-05-19T12:35:21Z

@wasp-codes
in the same folder as the .jpgs

MAJEEDI · 2020-05-19T21:03:49Z

HELLO, if you interest in this area i can send you sample from the annotation file.txt and image to show how can i read the images and draw the bounding box for object in the img. thanks in advance majeedi بتاريخ الثلاثاء، 19 أيار 2020 08:35:39 م غرينتش+8، Cindy Weng <[email protected]> كتب: Hey, might be a silly question but where do I put all the bounding boxes txt files? We mentioned the location of images but we did not mention the location of annotations to darknet did we? thanks in the same folder as the .jpgs — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

balajib363 · 2020-06-08T17:46:30Z

Hello,
I am trying to automate my annotations. I trained a model which can detect the object, so once an object is detected I am saving its x,y,w,h i.e(left_x: 43 top_y: 102 width: 498 height: 291).
But if I try to open this in Yolo_mark or labelImg I am not able to view it.
From this thread I understood I need to scale in between 0-1, but when I divide it with width or height its not near equal.

The below I got from LabelImg tool in Yolo format.
0.450000 0.515625 0.825000 0.610417

Please help how to get a relation between the detected box values and what yolo uses for training.

stevedepp · 2020-06-10T06:26:52Z

Hello, Thank you for the instructions. May I confirm? I am putting my custom data into darknet/build/darknet/x64/data/
Then, instructions say "To train on Linux use command: ./darknet detector train data/obj.data yolo-obj.cfg yolov4.conv.137" Are we running from ./darknet or from ~/darknet/build/darknet or from ~/darknet_custom/build/darknet/64 please ?

amashi01 · 2020-07-10T07:10:33Z

Are you able to get the formula? I am also having the same issue.

{'class_id': 0, 'width': 20, 'top': 387, 'height': 74, 'left': 789}, {'class_id': 1, 'width': 25, 'top': 348, 'height': 31, 'left': 805}, {'class_id': 2, 'width': 19, 'top': 447, 'height': 26, 'left': 826}, {'class_id': 4, 'width': 47, 'top': 545, 'height': 33, 'left': 727}, {'class_id': 3, 'width': 32, 'top': 364, 'height': 144, 'left': 896}, {'class_id': 5, 'width': 89, 'top': 246, 'height': 97, 'left': 825}, {'class_id': 7, 'width': 254, 'top': 224, 'height': 388, 'left': 725}

'image_size': [{'width': 1040, 'depth': 3, 'height': 780}]}

Emirismail · 2020-09-11T15:06:59Z

@sarratouil Could you please share your code?

saikrishnadas · 2020-10-07T06:43:50Z

Hey, I would like to know how to calculate .
This is how my data looks like.

With this i could easily find the widht and height but really stuck at finding the x,y that is need to convert to yolo format .

Help is appreciated :)

mk-hasan · 2020-10-07T08:06:48Z

Hi, You already have the bounding box information. That should be fine. Now you need to feed the data to yolo and check the code how it takes the data. Maybe I am not sure about your question, otherwise, I could help. Thank you.

kishore-jd · 2020-10-08T05:24:55Z

i am also facing the same issue

saikrishnadas · 2020-10-08T05:33:02Z

@mk-hasan Yolo takes in this format
My .csv file has these information expect . Yolo takes in center of x and center of y. My data has x_min ,x_max and y_min and y_max. Now how do i calculate the ?

kishore-jd · 2020-10-08T06:19:10Z

ForceQuell · 2020-10-22T19:51:21Z

Hi!
Could you tell how axes on image are oriented?
Do they oriented like in 2d-array (origin in left-top corner, X goes down, Y goes right)?
And does it all work for yolov4?

masoodazhar · 2020-10-27T18:03:22Z

.txt-file for each .jpg-image-file - in the same directory and with the same name, but with .txt-extension, and put to file: object number and object coordinates on this image, for each object in new line:

Where:

- integer number of object from 0 to (classes-1)
- float values relative to width and height of image, it can be equal from (0.0 to 1.0]
for example: = <absolute_x> / <image_width> or = <absolute_height> / <image_height>
for example:
<absolute_x> = (x minimum+<image_width>)/2
= <absolute_x> / <image_width> or = (image_height-1) / <image_height>

atention: - are center of rectangle (are not top-left corner)
For example for img1.jpg you will be created img1.txt containing:

stephanecharette · 2021-01-27T11:24:14Z

This old thread seems to come up a lot. I've added an entry to the FAQ with an example showing exactly how the numbers all fit together. See this: https://www.ccoderun.ca/programming/darknet_faq/#darknet_annotations

annezao · 2021-04-03T06:17:30Z

Is it a problem if somehow some values, after calculate to yolo format, look like this:

0 0.67890625 1.287037037037037 0.05364583333333333 0.2555555555555556

the y axis is above 1. Is that alright?

stephanecharette · 2021-04-03T08:46:57Z

If you're asking, I'm guessing you already know it is a problem. The values are normalized 0...1 so it is impossible to get a value > 1. And since that is the middle coordinate and not an edge of the rectangle, it should be impossible to get exactly 1.0 as well.

I'm not certain what that would do to Darknet during training, but it cannot be good. I wouldn't be surprised if it causes Darknet to crash as it attempts to create a RoI from the image outside of the image boundary.

muhammadhamzahabibhashmi · 2021-05-08T23:40:25Z

You have to follow a formula
x = (x1+x1+w1)/2
y = (y1+y1+h1)/2
with open('Annotations.txt', 'a') as file:
file.write(f'{classid} {y/512} {x/512} {(w1+20)/512} {(h1+20)/512}') # where 512 is the image width and height

stephanecharette · 2021-05-09T02:26:18Z

The format is described here in details: https://www.ccoderun.ca/programming/darknet_faq/#darknet_annotations

masoodazhar · 2021-05-10T17:42:59Z

yes. i have tried this. and this is working fine.
save detected coordinates like. if your image size is 512x470 or else.
just save the detected coordinates like example blow.

x,y,w,h = detected_coordinates # its corrdinates of a bounding box of an object of image
ymid, xmid, height, width = (((1+y1)/2)/y1, ((1+x1)/2)/x1, (y1-1)/y1, (x1-1)/x1)
print('{:0.6f} {:1.6f} {:2.6f} {:3.6f}'.format(xmid, ymid, width, height))

masoodazhar · 2021-05-10T17:44:35Z

Riankk123 · 2021-05-26T12:31:10Z

Does it mean that x_ mid = (left_x + width + left_x)/2 and y_mid = (top_y + height + top_y)/2 .
And bottom_y = top_y + height and right_x = left_x + width . ?
Please help ?

brandhsu · 2021-05-26T18:21:30Z

@Riankk123
Format is already stated here: #60 (comment)

AlexeyAB added the Solved The problem is solved using the correct settings label Jul 11, 2018

ycui123 closed this as completed Jul 11, 2018

JacobMarkBaird1998 mentioned this issue Aug 31, 2018

Exporting data in YOLO format NaturalIntelligence/imglab#83

Open

biparnakroy mentioned this issue Jul 16, 2020

How to auto create/export/produce annotation files for objects yolo detects on image? AlexeyAB/darknet#6240

Open

philipp-schmidt mentioned this issue Jan 9, 2021

Simplify gen_txt.py jkjung-avt/yolov4_crowdhuman#12

Closed

Flova mentioned this issue Jan 9, 2021

Issues on loading ground-truth bounding boxed eriklindernoren/PyTorch-YOLOv3#573

Closed

einareinarsson mentioned this issue Feb 22, 2021

What is yolo (.txt) format? rafaelpadilla/review_object_detection_metrics#15

Closed

pabsan-0 mentioned this issue Mar 1, 2021

YOLO darknet labels Unity-Technologies/com.unity.perception#226

Closed

nicOwlas mentioned this issue Aug 30, 2021

Add YOLO export format labelflow/labelflow#387

Closed

8 tasks

JamesButler10 mentioned this issue Mar 1, 2022

AU-AIR JSON to YOLO Conversion JamesButler10/FinalYearProject#3

Open

Arshadoid mentioned this issue Apr 5, 2023

Can't train instance segmentation on custom dataset ultralytics/ultralytics#1681

Closed

2 tasks

Specific format of annotation #60

Specific format of annotation #60

Comments

ycui123 commented Jul 2, 2018

AlexeyAB commented Jul 2, 2018

ycui123 commented Jul 2, 2018

ycui123 commented Jul 11, 2018

ycui123 commented Jul 11, 2018 • edited Loading

ycui123 commented Jul 11, 2018 • edited Loading

AlexeyAB commented Jul 11, 2018

ycui123 commented Jul 11, 2018

ycui123 commented Jul 11, 2018 • edited Loading

koutini commented Mar 12, 2019

sarratouil commented Mar 12, 2019

sarratouil commented Mar 12, 2019

koutini commented Mar 12, 2019

sarratouil commented Mar 12, 2019

AlexeyAB commented Mar 12, 2019

sarratouil commented Mar 12, 2019

sarratouil commented Mar 12, 2019

koutini commented Mar 12, 2019

sarratouil commented Mar 12, 2019 • edited Loading

koutini commented Mar 12, 2019

koutini commented Mar 12, 2019

AlexeyAB commented Mar 12, 2019

sarratouil commented Mar 13, 2019

sarratouil commented Mar 14, 2019

AlexeyAB commented Mar 14, 2019

koutini commented Mar 17, 2019

AlexeyAB commented Mar 17, 2019

koutini commented Mar 17, 2019

cindyweng commented May 19, 2020 • edited Loading

MAJEEDI commented May 19, 2020 via email

balajib363 commented Jun 8, 2020

stevedepp commented Jun 10, 2020

amashi01 commented Jul 10, 2020 • edited Loading

Emirismail commented Sep 11, 2020

saikrishnadas commented Oct 7, 2020

mk-hasan commented Oct 7, 2020

kishore-jd commented Oct 8, 2020

saikrishnadas commented Oct 8, 2020

kishore-jd commented Oct 8, 2020

ForceQuell commented Oct 22, 2020 • edited Loading

masoodazhar commented Oct 27, 2020 • edited Loading

stephanecharette commented Jan 27, 2021

annezao commented Apr 3, 2021 • edited Loading

stephanecharette commented Apr 3, 2021

muhammadhamzahabibhashmi commented May 8, 2021

stephanecharette commented May 9, 2021

masoodazhar commented May 10, 2021

masoodazhar commented May 10, 2021

Riankk123 commented May 26, 2021

brandhsu commented May 26, 2021

ycui123 commented Jul 11, 2018 •

edited

Loading

ycui123 commented Jul 11, 2018 •

edited

Loading

ycui123 commented Jul 11, 2018 •

edited

Loading

sarratouil commented Mar 12, 2019 •

edited

Loading

cindyweng commented May 19, 2020 •

edited

Loading

amashi01 commented Jul 10, 2020 •

edited

Loading

ForceQuell commented Oct 22, 2020 •

edited

Loading

masoodazhar commented Oct 27, 2020 •

edited

Loading

annezao commented Apr 3, 2021 •

edited

Loading