ryouchinsa / rectlabel-support Goto Github PK
View Code? Open in Web Editor NEWRectLabel is an offline image annotation tool for object detection and segmentation.
Home Page: https://rectlabel.com
RectLabel is an offline image annotation tool for object detection and segmentation.
Home Page: https://rectlabel.com
I may use EAST for a future project and I'll need to label non rectangular zones in image.
Do you think this type of labeling will fit in RectLabel?
Hello,
We have an automated process which detects salient features and automatically draws bounding boxes for us. We then open the images and corresponding Pascal VOC in RectLabel for final naming and cleanup of the labels. Because our saliency detection often find more objects of interest than we actually desire, I would like to be able to select and delete multiple annotations at once. Is it possible for you to add this feature?
Thanks,
Lonny Lundsten
Rectlabel can only convert to XML format and does not support conversion to JSON anymore. Also support for conversion to Kitti format for NVIDIA Digits would be great as well.
I often have to modify boxes copied with command + i from the previous image.
Usually I just have to move each boxes, add/delete one or two boxes and few resizes.
To select a box I have to right click it, the contextual menu pops in, left click outside the menu to discard it, click the center point to move the box, done.
Why can't I just left click the box to select it and move it?
Is there a better solution than right click?
Using the list of boxes on the right is not an option for me since I have many boxes with the same class name. I use this feature only when a bigger box do not let me select a smaller box that is inside the big one.
Sometimes in the train corpus, we can see the same image twice or an image that we do not want to label (since it has nothing to do here). So we need to delete it.
Now, I
I suggest to just :
This is a UX problem:
the labels box is stuck with the canvas, and they're centered in the app's window. If you make the window wider, only the sides go wider, but the size of the canvas is the same.
I suggest having the list of boxes stick to the right size, and the canvas (along with the "<", ">" buttons) take up all the remaining space. This makes it easy to still keep the whole document width when zooming in, for annotating small objects in big images.
As a plus, consider making the background of the app window a little bit darker in colour or putting a small border around the image canvas. This helps when you have images with white background (documents) and you cannot see where the image finishes and the rest starts. I admit my monitor is not perfectly calibrated and there may be a difference between the two, but it is too small.
Many thanks for your support !
Hi,
I'd like to ask you if it would be possible for me to download a previous version which is compatible with my OSX version (10.12)? As I've seen version 1.35 would be just fine for me.
Thanks,
Peter
It would be useful to have a prefix or re-locate tagged images to a different folder so i dont have to check which images are ready to train the model.
Would like to make some improvements:
specifically build in a translator so that the annotation files/file can better be transcoded to the different project bases - like tensor box..
Very willing to code improvements and drive Pos
Since I've upgraded to 1.35 the mode described in #12 in not working anymore.
Thank you for providing a nice tool.
RectLabel is useful for us to be able to load large size images.
But I found a problem that bboxes look like shifted to left side when RectLabel reads a Pascal VOC annotation file which is created with LabelImage. (https://github.com/tzutalin/labelImg)
Image Size: 10000 * 2500 pixels
Format: PNG
MacOS 10.13
Do you have any idea to fix it?
Thank you.
Is there support for negative coordinates (say, for an object with some occlusion)?
Hello, looks like the images are ordered as "339.jpg, 3390.jpg, 3391.jpg, ....". Is there a way I can order images by the numerical value instead of alphabetical order? I was trying to get images in the order "339.jpg, 340.jpg, 341.jpg ...". Thanks a lot!
When labeling a train data, we often label similar images in a row.
For example, I currently have a document type to label, I have many images of this kind of document, each image has nearly the same boxes with the same classes at roughly the same position.
It would be great to just have a hotkey to duplicate the boxes of the previous image and just adjust the positions.
Has a workaround I currently duplicate the JSON file and change the path of the image, but it's not very productive.
It is getting annoying and tiring to have the pop-up dialog for every annotation.
I believe it would be much much faster if we could select a label, then make all boxes for that label. No prompting in between. Then we change to the next label, and again do all boxes.
Obviously, if you have only 1 label, you never get the prompt, which is just getting in the wa
As I use RectLabel in a multi language environment, the default hotkeys are not always suitable for production use.
For example command + 0 (zero) to Zoom Fit needs two hands on a French keyboard, and going from mouse to keyboard and back is not very handy. If one can use his own hotkeys this would solve the problem.
I have personally implemented MASShortcut in my own application with success.
I am no longer able to open RectLabel without it quitting right as I open it, and then I get an error message asking if I should ignore, report, or reopen the app. I am on a Mac.
How can I resolve this issue?
We have an e-mail address to support RectLabel users.
[email protected]
Also we have a twitter account to inform new updates.
https://twitter.com/rectlabel
We had some feedback on e-mail and twitter.
"Awesome tool!! Any chance you could add templates to go into PASCAL VOC format or COCO, this would make it even easier."
https://twitter.com/EyeingChicago/status/858799956404183040
If you could find some problems for RectLabel, could you post the problem on this issues page?
We can share the problem and how to solve it among our users.
Best regards,
Ryo
Hi,
I'm trying to use rectlabel_create_coco_tf_record.py
to convert COCO to TFRecord.
Once I run the command, it tell me:
Traceback (most recent call last):
File "/Users/tester/codes/Rectlabel-support/rectlabel_create_coco_tf_record.py", line 195, in <module>
tf.app.run()
File "/Users/tester/codes/tensorflow_workspace/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 125, in run
_sys.exit(main(argv))
File "/Users/tester/codes/Rectlabel-support/rectlabel_create_coco_tf_record.py", line 184, in main
num_shards=1)
File "/Users/tester/codes/Rectlabel-support/rectlabel_create_coco_tf_record.py", line 160, in _create_tf_record_from_coco_annotations
annotations_list = annotations_index[image['id']]
KeyError: 1
When I trying to troubleshot for this error, I found the exported COCO json file not contained annotations
as below:
...
"annotations" : [
],
...
is that a expected behavior?
please advice me if anything else i forget to configure in RectLabel?
Thank you, it a awesome tool! :)
In the JSON file the boxes are saved in x, y, w, h
format.
An option to save the boxes as x1, y1, x2, y2
may help for some usages.
When I move the training set (Image directory and annotations subfolder) from a computer to an other and open this training set to the new computer, only 2 classes are present : face and shoes.
This occurs even if I can see the actual classes on the already labeled images.
Have I missed something?
Is the class list stored somewhere, so I can copy it?
Hey,
I noticed that the boundaries of annotations can exceed the image ranges, i.e.
I don't know if this is the desired behaviour, but I was definitely not expecting it.
It can happen in case I adjust my annotations and drag the anchors outside of the image boundary. I haven't tested JSON yet, but this definitely occurs with PASCAL-VOC format which broke some of my machine learning pipelines.
Version: 1.24
The file couldn't be opened.
file:///Users/.....
I'm getting this error message whenever I try to open a folder, same message whether it is a folder that I've already annotated in Rectlabel (with existing xml files) or a new one. Haven't used the app in a couple of months but was working well before
When trying to label objects, the bounding box is forcing me to use a square. What I mean is no matter how much I drag the mouse when I'm drawing the box or when I attempt to resize an already drawn box, it's forcing the aspect ration to be 1:1, so I can't draw a rectangle.
Rectlabel v1.26
Note: I did not see this behaviour on my 15" Macbook Pro with Rectlabel v1.24.
Working with tons of images in a dark room with a bright white app can really begin cause fatigue and eye strain. It would be nice to have a dark theme for working in 'dark room' type environments. Photoshop, Pixelmator, and Affinity Photo all have dark backgrounds and are excellent examples to emulate.
First of all, let me just say thanks for creating Rectlabel. I have found Rectlabel very simple and straight to the point whereas other labelling tools require PyQt4 which isn't compatible with modern macOS releases.
I find it quicker to just label the smaller objects right away than have to zoom in to get a clearer view of them. Resizing the windows to make them large causes some of the UIViews to disappear and the image view also gets displaced. I get around that by keeping the window size small enough for the UIViews to still be visible. It's not very high priority as everything else still works great but a few more Auto Layout constraints could definitely make this great app even better.
Hello,
each time I export csvs (convert to csv), the csv file contains different coords. I'm including a few different exports. There were no annotation changes between these saves.
I'm sure I'm overlooking something simple, but when I move to the next image in the directory the bounding box from the previous image already appears. Then in the XML, 'filename' is that of the next image, not the original.
Not sure how to save the bounding box for an image, then move to the next image, creating the next xml file in the process.
Hey!
Can't seem to save BBox file on MacOS Sierra, RectLabel downloaded from the app store. Error says a folder can't be created but doesn't give another option for saving the .json (in my case) file, nor does making my own annotations folder fix the problem.
Cheers,
Elizabth
Our corpus of train documents is growing quickly (yesterday I had 320 images, today I have 2396 images).
We need to seek at random position (to take a template of boxes for example) in the flow of images and with Command + arrow it is just not possible.
A search by index and by partial/whole filename would be great.
For example:
As title says it would be great to just use arrows while navigating.
I annotated a bunch of images and then realized that the setting was set to json which is not the format I want.
I switched to Pascal VOC but now I would have to re-annotate everything.
Is there a way to convert the json files to Pascal VOC xml?
Hello,
I can launch rectlabel form the terminal using this command: "open -a rectlabel"
Is there a way to pass in an argument to specify what folder to open? I have a need to launch rectlabel from a script, opening a different folder of images each time for labelers.
Thanks in advance,
Zach
There are a number of inconsistencies in XML and JSON formats that range from the treatment of image size:
XML =
4928
3280
JSON = "image_w_h" : [4928,3280],
to Objects:
XML =
carconenumber-1
2287
1865
2677
2099
JSON = "objects" : [
{
"label" : "carconenumber-1",
"x_y_w_h" : [2287, 1865, 391, 235]
}]
as well as challenges with respect to a concatenated annotation model supported for platforms like TensorBox.
The VOC format really expects:
/* tensorbox annotation format:
https://github.com/Russell91/TensorBox/blob/master/utils/annolist/readme.md
{
"images": [image],
"classes": [string]
}
image{
"id": int,
"image_path": string,
"rects": [rect]
}
rect{
"classID": int,
"x1": int,
"y1": int,
"x2": int,
"x2": int
}*/
so there may be some "custom" Tensorflow semantics here, but I'd love to create the encoders/decoders to support a multitude of different TF annotation structures?
open the base?
I made a video to talk about an idea I had for a simple program to generate small labeled images fast. You would load an image into the program and then use the mouse to take snapshots and automatically save them to folders for each class. You might use hotkeys to designate classes. Not sure if this exists already? https://www.youtube.com/watch?v=FDouW7fSIms
Awesome tool! But I think you need to support more export format options like COCO, JSON!
When I am creating a polygon outline for an object, sometimes I'll accidentally double click instead of single click.
In the current version of RectLabel, double clicking while drawing a polygon instantly deletes and clears the entire shape for some reason. This is very frustrating if you are annotating complex objects:
Is there any way to disable this feature? If not, could you consider that as a future feature?
Allow to enable an autosave function to save every annotation we make or modify automatically.
Selecting "unlock RectLabel" times out--just a spinning loading wheel. Sometimes a message appears saying to contact the App Store.
How do you purchase the program?
The app is only available for mac version 10.12 or above I have a 10.11 can I still use this?
Great app!
Would like it to export direct to Turi Create format. I know CSV export is implemented, but would be nice with a export for Turi button.
If you could do a smooth transition from pictures to Turi, I could easily see your app becoming a 'standard app' in the workflow of marking images for machine learning on macOS.
Thats all, have a nice day!
Would it be possible to automatically compute and add the minimum bounding box for a polygon object? This would be helpful for Tensorflow.
I can create boxes and they are shown in the right panel, but once I accept a box, it is no longer shown on the image. I tried to focus
or edit
or move
but it just won't show. So I have no idea what I am editing.
MacOS High Sierra
Anyway, your tool is the most convenient to use for OSX users, thank you!
It would be useful to map all the used atrributes to a label map, and why not generating directly the TFRecord files.
I use RectLabel to label images on my Mac. The resulting files are then copied to big server with plenty of CPUs and GPUs running under Linux.
The issue is that I need to modify each JSON file to match the file system of the Linux box.
If the path was relative between JSON and image file, my job would be easier ๐
A simple checkbox in preferences may do the trick.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. ๐๐๐
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.