How CVPR 2026 Papers Solve Motion Jitter, Pose‑Free Avatars, and Point Cloud Convolution

Natural Human Motion Recovery by Aligning High‑Order Temporal Dynamics from Monocular Videos

Recovering 3‑D human motion from monocular video often produces overly smooth or jittery motions because existing pipelines do not model higher‑order temporal signals such as joint velocity and acceleration. The paper identifies the lack of reliable high‑order dynamics as the root cause of unnatural motion.

HTD‑Refine addresses this by first applying PVA‑Net to jointly estimate 2‑D keypoints, 3‑D joint velocities, and 3‑D joint accelerations from the video. These estimates are then used as soft constraints that align the speed and acceleration of any existing human motion reconstruction (HMR) output, reducing jitter and excessive smoothing while preserving physical plausibility.

Evaluation on the challenging real‑world datasets RICH and EMDB shows consistent improvements in global trajectory accuracy and dynamic naturalness across multiple state‑of‑the‑art HMR methods, confirming the benefit of explicit high‑order temporal alignment.

UIKA – Fast Universal Head Avatar Modeling from Pose‑Free Images

UIKA is a feed‑forward, driveable Gaussian avatar model that can be constructed from an arbitrary number of pose‑free inputs, ranging from a single image to multi‑view captures or mobile video. The method introduces a UV‑guided pipeline: each input image is processed to obtain per‑pixel facial correspondence, which is re‑projected into a UV map that is independent of camera pose and facial expression.

Learnable UV Tokens allow attention mechanisms to operate simultaneously on screen‑space and UV‑space features. Aggregated UV information from all views is decoded into Gaussian attributes that can be driven by traditional skinning for real‑time rendering. The implementation achieves real‑time performance at 220 fps.

Quantitative experiments demonstrate that UIKA outperforms existing mainstream approaches in both single‑view and multi‑view configurations, delivering higher visual fidelity with substantially reduced modeling time.

PointCNN++ – Performant Convolution on Native Points

Current 3‑D point‑cloud convolution methods trade off between the high precision of point‑based approaches and the efficiency of voxel‑based ones. PointCNN++ extends sparse convolution from voxels to native points by formulating a point‑centered convolution as a matrix‑vector‑multiply‑reduce (MVMR) operation and implementing highly optimized custom GPU kernels.

Benchmarks indicate that PointCNN++ runs several times faster than traditional point‑based methods while reducing memory consumption by an order of magnitude. When used as a drop‑in replacement for voxel backbones, it improves point‑cloud registration accuracy and retains superior speed and memory characteristics.

Code and continuously updated releases are available at https://github.com/ant-research/pointelligence.