October 19, 2022
For each task group within a given application, it will be possible to attach (one or several) example photos and to consult them:
- Before taking the picture to avoid mistakes
- While taking the picture if any doubts
- After taking the pictures if there is a KO in the analysis
In order to increase the visibility on what is happening on the field, comments are now available:
- Comment can be left before taking a picture (if impossible to take the photo) or after the analysis of a photo in order to justify a KO
- Reasons are configurable for each application
- It is possible to force (or not) field workers to let a comment in case of a KO
With the purpose of promoting the annotation reviews, facilitating the review process, and reaching higher annotation conformity this feature enables to:
- Flag the annotation as being reviewed to insure the annotation double check
- Log review metadata as the last review date and the reviewer username
- Access to the gallery of annotated images that are waiting to be reviewed
- Switch between the concept's reviewed images and the not reviewed ones
- Manage the process by resetting existing reviews and arranging annotations for another round of reviews
Additional metrics at training time
New metrics are calculated and displayed during neural network training. These metrics allow you to better understand your training and especially overfitting, and thus achieve better performance.
Here are the validation metrics we display:
- For classification : we display cross-entropy loss and accuracy
- For Tagging: we display binary cross-entropy loss, accuracy and mean average precision (area under the curve of precision recall)
New blur detection algorithm for data conformity
In addition to duplicate photos within an application, you can now enable a blurred photo compliance check for your application, and thus reject photos that do not meet your compliance criteria.
- Exporting datasets by default generates files without ids
- The count of images in the trash is not anymore included in the project's total counts
- The date on which the annotation has been performed appears on the image information
- It is now possible to tag a batch of images as 'without concept'
- The audit tool in detection views is empowered by showing the concept's existing annotations along with the proposed errors
- The model version deployed no longer necessarily corresponds to the end of the training, but to the best epoch. For classification and tagging, during training, we save the model with the best validation metric (for Classification, we use Accuracy and for Tagging, we use Mean Average Prevision).
- It is possible to modify the
Work Order Typesof a work order after its creation. This change is automatically taken into account to adjust the list of associated tasks.
- Updating a site with a change of version of the deployed application is automatically taken into account, without any other operation required.
- Display the information (on work order page) regarding the number of tasks included in the
- The KPI
Required photosis now dynamically computed so we can see the evolution of config between 2 WOs
- Tasks are now displayed in the order defined in the application center
- The problem concerning the difficulty to scroll on android phones is now resolved
- Tasks in
push_to_field = falseare no more displayed on the analysis page
- When an annotation changes in the parent view, now it checks for the validity of the image, and its annotations in the child views
- 404 error when clicking on the model link in the app center is resolved
- The navigation problem between images in the 'Annotate images' section is resolved
- You can now use the add prediction button as an annotation feature on Studio
- The issue with the images which automatically annotated as 'without concept' is identified and fixed
- The issues relative to role assignments are resolved
- Audit images in children's view belong to a Detection parent view set to appear as cropped
- You can now change organizations via URLs without getting 404 errors