Maker Tool
Methods to Annotate
On the header are various methods to create annotations. Here are the following:
1. Box
It creates a rectangle for an object in the image. Hence determining the length and width of the object.
How to interpolate boxes created?
Interpolation helps in annotating objects present in sequential frames quickly.
The user should create an annotation in the frame where the object first occurs. This frame becomes the keyframe. Add attributes for the object in the 1st frame created
While the annotation is selected, navigate across a few frames to replicate the annotation in the proceeding frames.
Adjust the position of the annotation every few frames. The frames in which the annotation is adjusted becomes a keyframe. This means that all the replicated annotations between two keyframes will re-adjust by itself.
A keyframe can be identified by an underlined that appears on the seekbar for that particular frame.
It is advisable always to check the interpolated box between two keyframes and make any adjustments if needed.
Once the object is interpolated across all the frames it is present in, the user can check the attribute values.
The recommended method of labeling sequential images
Here is the method we recommend to our users to follow for increased efficiency:
Annotate the first object in the frame. Fill the attribute values for it.
Interpolate the object across all the frames it is seen in. Check the interpolated object between keyframes to see if there are any adjustments needed.
Since the attribute values get forwarded with the annotation to the next frame, check if the values need to be changed in the remaining frames.
Once the user has done the above, move to the next object and repeat the same.
Note that one can interpolate objects from 1st to the last frame or vice-versa. Hence it doesn't matter if you label a few objects from start to end of the sequence or the other way.
Why we recommend (A) - labelling an object across frames and then moving to label the next one and not (B) - Marking all the object in one frame and then forwarding them?
In method A, the user creates an annotation, interpolates it to the proceeding frames and checks for all its attribute values across frames, all at once before moving to annotate the next object. Hence there is lesser room for errors along with minimal time wastage.
Whereas in method B, if the user creates all the annotations in the 1st frame at the beginning, it will be hard to recall which have been forwarded to the proceeding frames (in cases where no. of objects to be labelled are very high). Hence the user will most likely have to keep navigating back and forth to check which annotation has not been forwarded; leading to wastage of few seconds every now and then.
In method A, since the user can create the next annotation from last to the 1st frame, there is no requirement of navigating back to the 1st frame for every annotation.
This is not the case in method B, since each time he will need to come back to the 1st frame to forward the annotation created to the proceeding frames.
2. Cuboid
It creates a 2D cuboid for an object in the image; determining the length, width, and depth of the object.
3. Polygon
It lets users create an enclosed shape around an object with straight connecting lines.
4. Line
It creates a line with a start and an endpoint. This is primarily used for lane markings.
How to segment edges of a line and give it attributes?
The 'Edge Labeling' feature allows the user to further segment a line with a given class into multiple segments where you can individually assign attributes to each segment.
Select the Line tool from the header and an appropriate class.
Once you start creating the line, you can simultaneously press the shortcut key
Ato mark the end of the first segment and the start of the next.Multiple segments can be marked on the line while creating it. Each of these segments can be given attributes, all being under the same class label given to the Line
5. Landmark
It lets users create a sequence of points to determine shapes and alignment.
How to Start Annotating?
Use Image Adjustment Controls to make the image quality clearer if required.
Create an annotation on the canvas using any of the methods mentioned above.
Select the class for it from the dropdown that appears.
The annotation detail panel slides open which has other class related attributes that need to be entered.
The annotation then appears on the Annotation List. All annotations created appear here along with details related to their attribute values and occurrence.
Here's a complete walkthrough of how to use the Maker Tool for 2D Image Labeling
How to use ML Proposals?
This feature proposes box annotations over the objects in an image using machine learning algorithms.
When you click 'Auto Label' (if this feature is active for the job), a slider appears where you can change the confidence score of the proposed boxes.
Higher confidence scores indicate higher accuracies of the proposed boxes. Therefore, when the scores are higher, the number of proposed boxes will be lesser.
Hence, the user can change and decide what confidence score works for them for the image they are labeling.
Last updated
