4 ms·

Show HN: Bounding-box labeler tool to generate the training data for YOLO v2

64 points by cartucho 9y ago

CardenB 9y ago
What makes this more appealing than a web interface? Presumably you still need a service to manage the data, and deploying to labelers will be difficult.
- cartucho 9y ago
  Thanks for the comment so the idea is to use OpenCV so that later it also supports video format and uses SIFT and Tracking OpenCV algorithms to make labeling easier. I wanted it to give automatic suggestions.
  - CardenB 9y ago
    Ohhh I see, cool
throwawaydl1 9y ago
This is a much better alternative and works completely in client side js http://www.robots.ox.ac.uk/~vgg/software/via/ http://www.robots.ox.ac.uk/~vgg/software/via/ http://www.robots.ox.ac.uk/~vgg/software/via/via_demo.html http://www.robots.ox.ac.uk/~vgg/software/via/via_demo.html
- cartucho 9y ago
  The idea is to use OpenCV so that later it uses SIFT and Tracking algorithms to make labeling easier. I wanted this tool to give us automatic suggestions for the labels!
bbernhard 9y ago
Interesting project, many thanks for sharing! On a side note: I am working on a client side annotation tool which is backed by OpenCV and Mask R-CNN. It is still in the early stages (and may contain some bugs), but in case you are interested, you can check it out here [1] [1] https://imagemonkey.io/annotate https://imagemonkey.io/annotate The whole service is running on a really small instance at the moment (will be moved soon), so it's rather slow at the moment.
manuaero 9y ago
Thanks for sharing. We've also been working on a similar tool (although does not output to YOLO format): https://www.labelbox.io/ https://www.labelbox.io/ We've open sourced the labeling interface for anyone to easily build and support labeling any data - as long as it can be loaded in a browser. Learn more here: https://github.com/Labelbox/Labelbox https://github.com/Labelbox/Labelbox About Labelbox: Labelbox is a enterprise grade and cloud based tool to easily label data for machine learning. Labelbox streamlines data labeling workflow, from micro labeling projects for quick R&D to production grade projects requiring hundreds of collaborators. It is agnostic to data type and has open source labeling frontend with already built templates for image classification & segmentation and text classification. One can label any other kind of datasets by creating a custom labeling interface with javascript API (labeling-api.js). Additional feature includes exporting data in JSON/CSV with auto generated image masks, project & team management and labeling analytics.
pronoiac 9y ago
I had to look it up: > "You only look once (YOLO) is a state-of-the-art, real-time object detection system." https://pjreddie.com/darknet/yolo/ https://pjreddie.com/darknet/yolo/ I'd also include the "image classifier" keyword.
- robotresearcher 9y ago
  It’s an object detector, not an image classifier.
- cartucho 9y ago
  Yes! There are very cool tutorials on how to train Yolo to detect your own objects: https://timebutt.github.io/static/how-to-train-yolov2-to-detect-custom-objects/ https://timebutt.github.io/static/how-to-train-yolov2-to-det...
- smnscu 9y ago
  You can go through a video frame-by-frame with YouTube, and it's worth doing it for the video on their landing page for moments such as: https://i.imgur.com/5tRzFsZ.jpg https://i.imgur.com/5tRzFsZ.jpg https://www.youtube.com/watch?v=VOC3huqHrss https://www.youtube.com/watch?v=VOC3huqHrss
infinitone 9y ago
I've used a bunch of these for object detection labeling.. the best one i found so far is labelmg (free) and rectlabel for Pascal VOC format.
pilooch 9y ago
Key to annotation is using pre trained or partially trained object detectors along with similarity search to quickly highlight unlabelled, similar objects in a dataset. Think of group labelling instead of labeling one object at a time. One of the open source implementations of object search out there is https://github.com/beniz/deepdetect/tree/master/demo/objsearch https://github.com/beniz/deepdetect/tree/master/demo/objsear... and that you can build your own group annotation tool around.
s-b 9y ago
This is really cool! Have you tried using superpixel segmentation? It works well for some biological applications. Grabcut is also immensely useful. We've been working on a platform for medical image and video annotation tasks. This means we need to support everything from DICOM to large pathology images, and endoscopy videos. We also have it connected to deep learning networks (e.g. Inception v3, YOLO, ENet) so you can easily train or download the JSON for offline analysis. We'd love to have beta testers from the HN community: https://semantic.md/annotate.html https://semantic.md/annotate.html
- cartucho 9y ago
  Thank you! I will give it a try with superpixel and grabcup!
ryouchinsa 9y ago
If you are using Mac OS X, you can use RectLabel. An image annotation tool to label images for bounding box object detection and segmentation. https://rectlabel.com https://rectlabel.com Key features: Drawing bounding box, polygon, and cubic bezier 1-click buttons make your labeling work faster Customize the label dialog to combine with attributes Settings for objects, attributes, hotkeys, and labeling fast Layer order for overlapped boxes Zoom in on a point Quick zoom to existing boxes Support the PASCAL VOC format

64 points by cartucho 9y ago

What makes this more appealing than a web interface? Presumably you still need a service to manage the data, and deploying to labelers will be difficult.

cartucho 9y ago
Thanks for the comment so the idea is to use OpenCV so that later it also supports video format and uses SIFT and Tracking OpenCV algorithms to make labeling easier. I wanted it to give automatic suggestions.
- CardenB 9y ago
  Ohhh I see, cool

throwawaydl1 9y ago

This is a much better alternative and works completely in client side js http://www.robots.ox.ac.uk/~vgg/software/via/ http://www.robots.ox.ac.uk/~vgg/software/via/ http://www.robots.ox.ac.uk/~vgg/software/via/via_demo.html http://www.robots.ox.ac.uk/~vgg/software/via/via_demo.html

cartucho 9y ago
The idea is to use OpenCV so that later it uses SIFT and Tracking algorithms to make labeling easier. I wanted this tool to give us automatic suggestions for the labels!

bbernhard 9y ago

Interesting project, many thanks for sharing! On a side note: I am working on a client side annotation tool which is backed by OpenCV and Mask R-CNN. It is still in the early stages (and may contain some bugs), but in case you are interested, you can check it out here [1] [1] https://imagemonkey.io/annotate https://imagemonkey.io/annotate The whole service is running on a really small instance at the moment (will be moved soon), so it's rather slow at the moment.

manuaero 9y ago

Thanks for sharing. We've also been working on a similar tool (although does not output to YOLO format): https://www.labelbox.io/ https://www.labelbox.io/ We've open sourced the labeling interface for anyone to easily build and support labeling any data - as long as it can be loaded in a browser. Learn more here: https://github.com/Labelbox/Labelbox https://github.com/Labelbox/Labelbox About Labelbox: Labelbox is a enterprise grade and cloud based tool to easily label data for machine learning. Labelbox streamlines data labeling workflow, from micro labeling projects for quick R&D to production grade projects requiring hundreds of collaborators. It is agnostic to data type and has open source labeling frontend with already built templates for image classification & segmentation and text classification. One can label any other kind of datasets by creating a custom labeling interface with javascript API (labeling-api.js). Additional feature includes exporting data in JSON/CSV with auto generated image masks, project & team management and labeling analytics.

pronoiac 9y ago

I had to look it up: > "You only look once (YOLO) is a state-of-the-art, real-time object detection system." https://pjreddie.com/darknet/yolo/ https://pjreddie.com/darknet/yolo/ I'd also include the "image classifier" keyword.

robotresearcher 9y ago
It’s an object detector, not an image classifier.
cartucho 9y ago
Yes! There are very cool tutorials on how to train Yolo to detect your own objects: https://timebutt.github.io/static/how-to-train-yolov2-to-detect-custom-objects/ https://timebutt.github.io/static/how-to-train-yolov2-to-det...
smnscu 9y ago
You can go through a video frame-by-frame with YouTube, and it's worth doing it for the video on their landing page for moments such as: https://i.imgur.com/5tRzFsZ.jpg https://i.imgur.com/5tRzFsZ.jpg https://www.youtube.com/watch?v=VOC3huqHrss https://www.youtube.com/watch?v=VOC3huqHrss

infinitone 9y ago

I've used a bunch of these for object detection labeling.. the best one i found so far is labelmg (free) and rectlabel for Pascal VOC format.

pilooch 9y ago

Key to annotation is using pre trained or partially trained object detectors along with similarity search to quickly highlight unlabelled, similar objects in a dataset. Think of group labelling instead of labeling one object at a time. One of the open source implementations of object search out there is https://github.com/beniz/deepdetect/tree/master/demo/objsearch https://github.com/beniz/deepdetect/tree/master/demo/objsear... and that you can build your own group annotation tool around.

s-b 9y ago

This is really cool! Have you tried using superpixel segmentation? It works well for some biological applications. Grabcut is also immensely useful. We've been working on a platform for medical image and video annotation tasks. This means we need to support everything from DICOM to large pathology images, and endoscopy videos. We also have it connected to deep learning networks (e.g. Inception v3, YOLO, ENet) so you can easily train or download the JSON for offline analysis. We'd love to have beta testers from the HN community: https://semantic.md/annotate.html https://semantic.md/annotate.html

cartucho 9y ago
Thank you! I will give it a try with superpixel and grabcup!

ryouchinsa 9y ago

If you are using Mac OS X, you can use RectLabel. An image annotation tool to label images for bounding box object detection and segmentation. https://rectlabel.com https://rectlabel.com Key features: Drawing bounding box, polygon, and cubic bezier 1-click buttons make your labeling work faster Customize the label dialog to combine with attributes Settings for objects, attributes, hotkeys, and labeling fast Layer order for overlapped boxes Zoom in on a point Quick zoom to existing boxes Support the PASCAL VOC format