FRAME: Floor-aligned Representation for Avatar Motion from Egocentric Video

TL;DR

This work introduces a lightweight VR-based data collection setup with on-board, real-time 6D pose tracking with a novel training strategy to enhance the model’s generalization capabilities and proposes FRAME, a simple yet effective architecture that combines device pose and camera feeds for state-of-the-art body pose prediction.

Abstract

Egocentric motion capture with a head-mounted body-facing stereo camera is crucial for VR and AR applications but presents significant challenges such as heavy occlusions and limited annotated real-world data. Existing methods rely on synthetic pretraining and struggle to generate smooth and accurate predictions in real-world settings, particularly for lower limbs. Our work addresses these limitations by introducing a lightweight VR-based data collection setup with on-board, real-time 6D pose tracking. Using this setup, we collected the most extensive real-world dataset for ego-facing ego-mounted cameras to date in size and motion variability. Effectively integrating this multimodal input -- device pose and camera feeds -- is challenging due to the differing characteristics of each data source. To address this, we propose FRAME, a simple yet effective architecture that combines device pose and camera feeds for state-of-the-art body pose prediction through geometrically sound multimodal integration and can run at 300 FPS on modern hardware. Lastly, we showcase a novel training strategy to enhance the model's generalization capabilities. Our approach exploits the problem's geometric properties, yielding high-quality motion capture free from common artifacts in prior works. Qualitative and quantitative evaluations, along with extensive comparisons, demonstrate the effectiveness of our method. Data, code, and CAD designs will be available at https://vcai.mpi-inf.mpg.de/projects/FRAME/

Figures (6)

• Figure 1: We introduce a large scale egocentric dataset (b) collected with a custom-made wearable capture rig (a). With this data we train FRAME, which takes as input a series of egocentric views and the VR pose tracking and predicts the skeletal motion of the user (c).
• Figure 2: Comparison of collection devices. With (a), (b), (c) an additional checkerboard must be mounted on top to align ground truth labels, making the helmet significantly top-heavy. Our device can be used for longer period of times and has a camera positioning that mimicks the most a realistic VR scenario.
• Figure 7: $\mathcal{L}$ and $\mathcal{R}$ are the left/right camera frames, $\mathcal{M}$ is the middle frame computed as the average of the two camera frames. The $x, y, z$ axes are color-coded to red, green, and blue, respectively. $\mathcal{F}$ is obtained by moving the origin of $\mathcal{M}$ at the ground level, aligning $y$ axis to be vertical, and using the projection of the $x$ axis of $\mathcal{M}$ on the horizontal plane to determine the direction of the horizontal axes of $\mathcal{F}$
• Figure 8: Visualization of the k-fold Cross Training Caching Strategy with $k=3$. $H_i$ denotes the hold-out data for the $i$-th run.
• Figure 9: Qualitative comparison on challenging inputs. The predicted 3D poses are overlayed onto external reference views not used for tracking. Our qualitative results confirm that our method predicts more accurate body poses, and significantly better handles contacts with the floor and lower limbs compared to prior state-of-the-art approaches unrealegoegoglassegoposeformer.