Labeling 3D cuboids in point cloud sequences can be a tedious job. Our batch mode and keyframe interpolation features are already of great assistance, but today we go one step further.
Smart cuboid propagation
Our new smart cuboid propagation feature – previously in closed beta – is now available to all users. Just draw a cuboid in one frame, open the batch mode, and click the propagation button: the cuboid will automatically be propagated to the next frames through automated tracking, making it much faster to label moving objects. If you see the tracking go off the rails, you can always pause the propagation, make a manual correction, and continue from there.
We’ll be improving this feature continuously, but note that for now, it works best with relatively dense point clouds.
Support for fisheye camera and other distortion models
Many robots and self-driving vehicles are equipped with fisheye cameras, which have an ultra-wide-angle lens. Fisheye cameras are useful for their wide field of view, but as a side effect, they produce images with strong visual distortion where straight lines are mapped to curved ones.
Most web-based visualization and labeling tools don’t support this type of images natively and instead require users to “undistort” their images before uploading them.
We’re proud to announce that we’ve tackled the hairy technical problem of bringing native support for fisheye cameras to our tools. This means that you can just upload your original fisheye images along with the camera calibration parameters, and see your point cloud points and annotations overlaid on top of them.
Besides the fisheye distortion model, we also support Brown-Conrady distortion.
We’ve optimized the performance of the point cloud cuboid interface for scenes with many cuboids. Whereas previously, the interface could become slow and laggy as the number of cuboid annotations increased, it now stays running smoothly even on older hardware.
Other features and improvements
The objects sidebar in the labeling interfaces has a new settings tab where you can change visualization, object coloring, and hotkey settings.
You can now add an is_static attribute to categories that can be static (e.g. a car, which can be moving or not). When this attribute is checked, the position and dimensions of that cuboid will automatically stay synced across frames.
The insights tab has a new subtab where you can see and export per-sample metrics like label time, review time, and the number of times a sample was rejected.
In the point cloud interface, the thumbnail image corresponding to the currently visible camera image in the synced image viewer is indicated with a small dot. This makes it easier to quickly open that image when closer inspection is needed.
For sequence data, you can now upload frame timestamps. The keyframe interpolation feature and kinematics feature take these timestamps into account when present.