About the series
As you know, we at Humans in the Loop have a great love and appreciation of a well-designed annotation tool. After the great feedback on the reviews we published of our the best platforms on the market here and here, we decided that it’s time for a deep dive in some of our all-time favorites!
This article is the last one from a series of 10 reviews which were published each week. The full list of 10 articles can be found on our blog.
The whole series is based on the premise of transparency and honesty and none of these reviews are sponsored. They are just our way to give props to the best teams out there working on making annotation easier for AI teams, and to share some of the know-how that we have been accumulating over the past few years as a professional annotation company.
As in previous reviews, our parameters are:
- project management
If you have additional questions or want to get in touch with us to beta test or feature your tool in an upcoming article, feel free to email us at email@example.com!
Understand.ai is a platform for training and validation data specialized for autonomous driving. Headquartered in Germany – the heart of the European automotive industry, the platform was created in 2017 with the goal of automating data annotation through a human-machine combined approach. The startup was acquired by the electric mobility company dSPACE in July 2019.
The “Zero-Touch Annotations” technology is what the team is most proud of and which we will be reviewing today. In addition, they offer some awesome solutions for anonymization and scenario simulation for autonomous driving – which will not be covered but which you should definitely check out.
Understand.ai offers a fully-managed solution for accelerating the training and testing of autonomous driving solutions, including technical support by their team with setting up a project, uploading data, etc. Pricing is available on demand.
The most exciting thing about Understand.ai is the fact that its annotation interface is entirely suited for autonomous vehicle annotation and it supports both 2D and 3D data. In fact, it’s the only tool on our series of 10 reviews which has a sophisticated LiDAR annotation interface for sequential data.
The 2D labeling interface supports bounding boxes, polylines, and polygons both on images and on video (note that bounding boxes and polygons cannot be used in the same task yet). The LiDAR 3D point cloud interface supports cuboids. Pixel-wise and voxel-wise (in 3D) segmentation can also be enabled on demand.
In 3D annotation you get access to helpful analytics of the cuboids, such as the distance to the camera, the height, width and length in meters, as well as the pitch, roll and yaw. When editing the cuboids, one can choose between several modes (‘translate’, ‘scale’ and ‘rotate’).
Arguably the best feature in this interface is that users can visualize and switch between different views of the camera (e.g. seeing the front, right, and top view of the bounding box, as well as the LiDAR and visible light camera scenes). This helps enormously to increase the accuracy and speed of annotation.
One of the best benefits of the platform in terms of project management is the assumption that labeling is a continuous process with many different steps involved. Projects on Understand.ai are managed as ‘Workflows’ within the ‘Taskmaster’ interface.
For each workflow, project managers can set up a certain sequence of steps, such as: Split (e.g. ‘split the video in chunks of 10 frames’), Annotate, Check, Merge (e.g. ‘merge the chunks after annotation’), and Interpolate. Some of these can be set up to run automatically whenever new data is available. Annotation steps can also be split based on the annotation type (e.g. first bounding boxes and then polygons) or based on the class (e.g. annotating vehicles every 10 frames while annotating pedestrians every 3 frames for higher precision).
Project managers can also track how data is moving through the steps by viewing how many tasks are “ready”, “in progress”, “done” or “failed” at each step. Tasks can also be assigned a certain priority so that they appear first in the labeling queue. At ‘Check’ stage, supervisors can submit feedback on a very granular object-level by pinpointing the specific issue.
Because of the tight specialization of the Understand.ai platform, they have been able to achieve great results in their efforts of automating the labeling of sequential data for autonomous vehicles in both 2D and 3D. The automation currently comes in two features.
The first one of these is propagation of one or all the annotations on the following frame. This applies to both 2D and 3D labeling and all objects preserve their parameters such as size and attributes. The key here is that once the propagated objects are corrected by the user on frame 2, the propagation on frame 3 assumes the same difference in position as between frames 1 and 2. Essentially it is based on the velocity of the objects and becomes a much better fit frame after frame.
The second type of automation is interpolation between keyframes. During the ‘Split’ step the user can decide how many keyframes will be annotated (e.g. 1 of every 5 frames) and during ‘Interpolate” step the rest of the frames are automatically labeled using linear interpolation.
Currently, Understand.ai is developing the NextGen Automation which will include pre-labeling of 2D and 3D data using a generic Neural Network. This will provide automatic detection of vehicles, as well as pedestrians and cyclists. This will make manual human labor necessary only for the quality checking and verification of the data, so we are very excited about the upcoming developments in 2021!
Hope this was helpful! If you are working on an AI project and are currently reviewing which tool might be the most appropriate for it, get in touch with us and we would be happy to have a call and advise you on the best way to build your pipeline.