1. Bounding Box Software.
2. 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.

I want to create my own training dataset for object detection in image. Does anyone know some good image labeling tools?

p.s. Labelme is not suitable for my case as it is web-based and the data uploaded will be in public.

Thanks in advance!

Comments

RectLabel An image annotation tool to label images for bounding box object detection and segmentation. 'RectLabel - One-time payment' is a paid up-front version.

you could try the one from opencv/apps

you could try the one from dlibit's in dlib/tools/imglab

Sparco Labels Template

Like being said by @berak, OpenCV got it's own tool for some months now. If you have suggestions for improvements, please sent back some feedback!

Trusted Mac download RectLabel for object detection 3.03.8. Virus-free and 100% clean download. Get RectLabel for object detection alternative downloads. Lookuq Lens - Object Detection and Recognition. Tiny Object Detection.

Hi, any document about how to use it ? Thanks

It is quite straightforward. Supply it with a folder of images and an output location of the text file. It will then open up and allow you to annotate regions. I have yet to build documentation for it, havent got the time yet due to research project deadlines.

Hello, try this one;

http://alpslabel.wordpress.com

This document is a walkthrough of testing the DeepForest prebuilt model for a new set of annotations.

Goal¶

For this example, we would like to test the accuracy of the prebuilt model for a section of Yellowstone National Park. Data for this example can be downloaded from the NEON portal under site code YELL, along with hundreds of other tiles from the same site.

Evaluation Data¶

Object detection models generally require that all target objects be labeled in an image. This means that we should cut the evaluation/training tiles into sections suitable for hand annotation. For this example, these crops will be small.

After about 25 minutes of annotations, the evaluation data is complete.

The data are currently stored in an xml format. These need to be converted to the standard keras-retinanet format with one bounding box on each line. For annotations made in RectLabel, DeepForest has a parse function preprocess.xml_to_annotations. For all other annotations type, the final output must match the following csv format, including column names.

Evaluation windows¶

Often the evaluation tile is too large to be predicted as single image, due to both the memory constraints and the size of the image objects. The prebuilt model was created using square crops of 400px. Therefore we need to use the preprocess.split_raster function to perform the following operations.

Arcgis Label Features

1. Split the evaluation tile into windows of 400px with an overlap of 5% among windows.
2. Save the windows as new images with unique filenames.
3. Gather the correct annotations that match each window.

Evaluate the prebuilt model¶

Before training a new model, it is helpful to know the performance of the current benchmark model.

Evaluate prebuilt model. We can view predictions by supplying a save dir (”.” = current directory). Predictions in green, annotations in black.

returns

These are pretty strong results, likely because the images are similar to those used to train the prebuilt model. In our experience, scores over 0.5 are unlikely to improve without significant additional training data, targeting a specific situation in which the model is performing poorly.Learn more about the mAP metric.

Rectlabel

Get a sample prediction¶

To view a prediction from the model, use predict_image.

Rectlabel Github

Predict a large extent¶

Rectlabel Ubuntu

To predict a large extent, we can crop overlapping windows, predict trees for each window, and reassemble the final tile after applying non-max suppression to the trees that overlap in multiple windows. Non-max suppression finds overlapping boxes and keeps the box with the higher confidence score. The threshold of overlap can be set using the iou_threshold argument.