Summary: You asked, we answered. We expended serious effort adding valuable features that we believe make the Theia user experience that much better. Thank you for all the excellent recommendations.
In general, I don’t typically write about releases or their features. It doesn’t really fit the mold of this blog, and honestly, changes between releases can be subtle; something that you may never see or encounter that took a month of development work to change. So, it never really appealed to me to bore this audience with these technical discussions. For us though, Theia 2023 is a little different.
Over the last many months, we have been working hard, not only from the typical data side of things (and improving machine learning algorithms), but also from a feature perspective. The focus of this blog is to give a summary of these features, and how they may influence your use of Theia 2023 (See the accuracy blog for a run down on how accuracy has been improved in this version).
For starters, we reorganized the menu bars to be more consistent with their actions:
Figure 1: Newly arranged menu bar at top left.
Other notable additions include changes to: Calibration, Tools, and Settings.
For Calibration:
We have added a ‘check calibration’ method, which will assess whether a camera has been bumped, or improperly calibrated to begin with.
Under the chessboard calibration, we have added the option to load intrinsic parameters in that pane (so no flipping back to the intrinsics pane to load them), use two cameras instead of three for origin triangulation, and allow changes to the long axis of the checkerboard, the normal direction, as well as the offset.
Under the object calibration, we have added the ability to manually adjust the origin. In some cases, it’s difficult to detect the chessboard in the final frame, and the frame needs to be propped up, or a frame in the middle of the trial needs to be used instead. This feature removes this requirement. After the chessboard calibration is finished, even if the origin is located at a previous and undesired frame, the user can input the four corners of the board and manually select these locations. After clicking them in three views, clicking adjust reference frame will change the reference frame to that location, while preserving the relative position and orientation of the cameras. So, you can get the accuracy of the chessboard calibration, with the flexibility of the object calibration.
Figure 2: Chessboard calibration improvements.
With these changes (and the image enhancement described later), there should be very few environments where the calibration fails.
For Tools:
Added an image enhancement pane to help clean up video quality that is less than optimal. For calibration especially, this can prove very useful to adjust the color of the videos to better reflect what it should be seeing (dark videos for instance). All of these enhancements can also be used when running analysis!
Added a ‘check synchronization’ tool that can assess the quality of the synchronization between the videos. If you think one of the cameras is off for any reason, this check can help you quantify that.
Re-designed the ‘assign calibration files’ tool to allow for more granular selection.
Added modify people IDs, which allows you to re-order the people after analysis. For instance, if you want person 0 to be person 1, that is an easy change. Also allows you to delete tracked people if you like, such as people in the background. In patch 1 of this release, we will be introducing a merge option.
Added an option to toggle views, which allows you to deactivate a particular view so that it is not used during analysis. This can be helpful if one of your views was bumped and the calibration is a bit off (found using the check calibration command…).
Figure 3: New video enhancement tools.
In addition to the changes described above, there are many small usability changes that should just make your life better; like launch speed being about 5x faster, and more complete error messages after any failure.
The batching application has been completely redesigned. The new design allows:
Analyzing a subset of trials:
Regular expression filtering to sort the trials
Checkboxes to select and deselect specific trials or groups of trials
Showing the files that are about to be analyzed, so you have a better understanding of the trials that are found by the batch app.
Automatically adding default preferences, and allowing you to change and apply these preferences to subsets of trials more easily.
Verifying that the data structure is set up correctly with preferences and calibration files, flagging incomplete data before the batch has started!
Figure 4: The re-designed batching application in use.
Importantly, these changes, as well as every other feature within the application have been comprehensively detailed in a series of tutorial videos (the videos are unlisted on our channel but links are available via the download portal), so you should have no problem learning how to use them.
Last, but absolutely not least, we have increased the number of keypoints that we track (now 124), and the overall size of the data set significantly. These changes, and how they influence the model performance, are captured in the accuracy blog of Theia 2023.
So yes, a lot of good things are going on! Our goal is to provide software that is as accurate and as usable as possible for our customers. If this is something that resonates with you, send us an email and we can provide you with more information!
To learn more about Theia, click here to book a demo.
Kommentare