updated: November 28, 2024

On this page we showcase the performance of our latest Avatars from Photos (AFP) ****model. This page is designed to provide an in-depth look at how our model outperforms its predecessor and competes with leading solutions in the market. Through detailed benchmarks across three diverse datasets, we aim to give a clear understanding of the model’s enhanced accuracy, versatility, and reliability.

<aside> ➡️

TLDR: Benchmark Summary

The benchmark results highlight significant advancements in our Avatars from Photo 2.0 model, showcasing improved accuracy across all metrics: PA-MPJPE, MPJPE, and PVE, compared to AFP1.0.

1. Pose Accuracy (PA-MPJPE): AFP2.0 achieves a substantial improvement in PA-MPJPE, with reductions of 25.2% on EMDB and 26.4% on RICH, demonstrating its enhanced ability to accurately align poses in both dynamic and diverse body scenarios.
2. Joint Positioning (MPJPE): With reductions of 23.7% on EMDB and 11% on RICH, AFP2.0 significantly improves joint positioning accuracy, ensuring more precise pose estimation for challenging and varied datasets.
3. Combined Pose and Shape (PVE): AFP2.0 reduces PVE by 24.5% on EMDB and 12.1% on RICH, indicating its superior capability to capture both pose and shape for highly dynamic and inclusive scenarios.

These results affirm that AFP2.0 delivers measurable and impactful improvements in pose and shape accuracy compared to leading methods, making it a robust solution for diverse real-world and expressive applications.

</aside>

Benchmark Details

Datasets Used

Metrics Evaluated

Benchmark Results

3D Pose & Shape Evaluation on EMDB

3D Pose & Shape Evaluation on 3DPW

3D Pose & Shape Evaluation on RICH