Summary
Theia3D's 2024 update, Apollo, is fast approaching! Please stay tuned over the next two weeks as we will be releasing multiple blog posts detailing some of the improvements to look forward to later this month! In this blog post, we provide an update on the validation of Theia3D Apollo against Theia 2023 and our reference marker dataset. This includes an examination of numerous biomechanical measures such as segment and joint angles, and shows subtle but meaningful improvements in several measures across multiple movement types, including slow, fast, cyclic, non-cyclic, bilateral, and unilateral movements. For interested readers, we recommend keeping an eye on the blog for more Theia3D Apollo update posts including a discussion of model changes and the effects of varying degrees of freedom on Theia markerless tracking. Stay tuned!
At Theia, we are committed to continuous improvement and transparent documentation of the validation process we use to evaluate our markerless model. Therefore, to prepare you for the upcoming release of Theia3D Apollo, we’re providing you with an updated report on how this year’s model compares against our reference marker-based dataset, as well as previous versions of our software.
With last year’s release of Theia3D 2023, we saw subtle but meaningful improvements in our pose estimation and reduced signal variability across a variety of movements, especially during challenging poses. This year, it’s not surprising that we see similar changes and improvements. At some point we may have to address the long-term expectations for changes to our pose estimates, as we’re already beginning to see evidence of asymptotic improvement - but we’ll leave this for a separate discussion. Once again, we’re using our validation dataset containing several movements: walking, running, countermovement jumps, double-legged drop jumps, single-legged drop jumps, and squats. This dataset allows an assessment of Theia3D’s tracking for different scenarios and movement types: cyclic, non-cyclic, slow, fast, bilateral, unilateral, etc. Given the numerous types of results and the overall volume of data, we’ve summarized the main observations for each movement type below.
Walking
As one of the most common movements studied in the field of biomechanics, walking gait is typically our first and main validation movement of interest.
Below we have presented the segment angles for the torso, pelvis, and lower limbs from our marker-based reference dataset, Theia 2023 (v2023-1-0-3160), and Theia3D Apollo (v2024-1-0-4409). To summarize the changes, we see that the torso and pelvis segment anterior tilt angles (X angles) are now much closer between the marker reference data and Theia3D Apollo, compared to Theia 2023. In a future blog post we'll discuss the model changes in detail and how the Theia3D Apollo hip joint centers are consistently narrower than both marker and Theia 2023 which more accurately reflects true hip joint widths based on imaging studies; this model change results in the slight shift seen in the thigh segment Y-angle. The foot segment for Theia3D Apollo is now slightly more supinated and externally rotated compared to the marker data and Theia 2023, a change which will also be addressed in a future post. Rather than representing any meaningful difference in how the foot segment is tracked, these differences show how very small changes to segment definitions can easily cause small shifts in segment angles. Across most of the remaining segment angles, Theia3D Apollo continues to demonstrate excellent tracking and there is little improvement - for example, in the foot, shank and thigh X-angles, and thigh and shank Y-angles. The most challenging segment angles continue to be those that capture axial rotation of the shank and thigh, which has changed little between the two Theia versions shown here, but these angles continue to show relatively less variability compared to marker-based measures. Overall, the only segment angle that has an RMSD relative to marker data that is greater than 5 degrees are the thigh and shank Z-angles.
Figure 1: Segment angles from the reference marker dataset (black), Theia 2023 (teal), and Theia3D Apollo (blue) for the torso, pelvis, thigh, shank, and foot segments during walking.
When we move on from individual segment poses and review the results of the lower limb joint angles, we see that there is no longer any offset in the hip joint angles, which nearly perfectly match those from the marker dataset. The hip ab/adduction joint angles are similarly affected by the narrower hip joint centers, with a slightly reduced joint angle value during early stance and late swing. The sagittal plane ankle and knee joint angles continue to show excellent agreement with those from the marker data, with a very slight shift toward plantarflexion in the ankle angles during stance but closer agreement during swing. As expected based on the modified foot segment definition, the ankle ab/adduction and internal/external rotation angles have similarly shifted towards adduction and external rotation. And of course, the hip and knee internal/external rotation angles continue to capture joint angles with considerably lower variability than marker data, and a distinctly different pattern in hip internal/external rotation. Overall, all of the sagittal and frontal plane joint angles have RMSDs of less than 6 degrees, while the internal/external rotation angle RMSDs are 9 degrees or less.
Figure 2: Joint angles from the reference marker dataset (black), Theia 2023 (teal), and Theia3D Apollo (blue) for the ankle, knee, and hip during walking.
Running
Alongside walking gait, running is one of the most studied movements in biomechanics and is therefore an important validation task that represents a faster movement than walking. As one might expect based on the results from walking, running kinematics from Theia3D Apollo showed similar changes relative to marker and Theia 2023 results. Based on the segment angle results, we see that Theia3D Apollo tracks the torso more similarly to the marker data than Theia 2023, now with only a barely perceptible offset in the sagittal plane and a very slight reduction in range of motion in the frontal plane. Several observations from the walking results also apply here, including the reduced offset in the pelvis anterior tilt, shifted frontal plane thigh angles due to the narrower hip joint centers, and foot segment angles that are shifted slightly towards supination and external rotation due to the adjusted foot definition.
Figure 3: Segment angles from the reference marker dataset (black), Theia 2023 (teal), and Theia3D Apollo (blue) for the torso, pelvis, thigh, shank, and foot segments during running.
The changes in running joint angles were also consistent with those observed in walking, including the reduced offset in hip flexion angle relative to the marker data, hip ab/adduction angles with reduced adduction due to the narrower hip joint centers, and ankle joint angles with shifts towards supination and external rotation.
Figure 4: Joint angles from the reference marker dataset (black), Theia 2023 (teal), and Theia3D Apollo (blue) for the ankle, knee, and hip during running.
Countermovement Jumps & Squats
Countermovement jumps are a dynamic, non-cyclic movement with a large lower limb range of motion and provide a different but useful perspective as a validation movement. Squats are similar and can include an even greater range of motion with more deliberate, less dynamic movement. For the sake of brevity, we’ve discussed the results from the countermovement jumps here, but the same observations can be made based on the results of the squats.
From the results below, we see many similarities with the previous two tasks, including sagittal plane torso, pelvis, and hip joint angles with reduced or eliminated offsets, frontal plane thigh and hip angles shifted away from adduction due to the narrower hip joint centers, and foot and ankle joint angles that capture slightly more supination and external rotation. Theia3D Apollo measures lower range of motion in axial rotations of the thigh and shank segments, and their corresponding joint angles, as well as all foot segment angles and their associated ankle joint angles. Some of these changes may represent reduced sensitivity to skin-attached markers and their tendency to capture axial rotations, for instance those movements of the shank and thigh during the high-flexion phases of the movement (countermovement, landing), when the measured axial rotations are more likely to be from joint angle crosstalk than actual joint movement.
Figure 5: Segment angles from the reference marker dataset (black), Theia 2023 (teal), and Theia3D Apollo (blue) for the torso, pelvis, thigh, shank, and foot segments during countermovement jumps.
Figure 6: Joint angles from the reference marker dataset (black), Theia 2023 (teal), and Theia3D Apollo (blue) for the ankle, knee, and hip during countermovement jumps.
Drop Jumps
For one last movement assessment, we include bilateral and unilateral drop jumps, which are another dynamic and non-cyclic movement with even greater accelerations than countermovement jumps. Once again, we see similar findings across both types of drop jump (double-legged and single-legged), including reduced or eliminated offsets in sagittal plane torso, pelvis, and hip angle measures, slightly shifted frontal plane thigh and hip angles due to the narrower hip joint centers, and slightly shifted foot and ankle supination and internal/external rotation angles due to the modified foot segment definition. The remaining measures show minimal changes relative to Theia 2023, and generally provide measurements that are consistent with the marker data, although some measures exhibit offsets (e.g. pelvis and thigh Z angles, hip internal/external rotation angles). For the double-legged drop jumps, across all segment angles only the shank Z angle RMSD was greater than 5 degrees, and across all joint angles only the knee internal/external rotation RMSD was greater than 6 degrees. For the left and right single-legged drop jumps, patterns among the segment and joint angle RMSDs were consistent for both sides; only the thigh and shank Z angles RMSDs were greater than 6 degrees, while the frontal plane knee and ankle joint angles and transverse plane angles at all three joints had RMSDs between 5-10 degrees.
Conclusion
Overall, the Theia3D Apollo update provides small but meaningful improvements in whole-body 3D pose estimates compared to Theia 2023. Particular improvements include the torso and pelvis segment orientations, and hip joint center positions and associated thigh and hip angles. The remaining measures demonstrate consistent or slight improvements in tracking compared to previous versions of Theia3D and relative to marker data. These changes and improvements are seen across numerous movement types, for fast, slow, cyclic, non-cyclic, bilateral, and unilateral movements, supporting the continuous improvement of our markerless model for generalized movement tracking.
Check out our future posts on Model Definitions and Degrees of Freedom for more info on Theia3D’s upcoming Apollo update. And for more info or to book a demo, you can reach out to us here.