Research map / 3D vision / Human pose estimation

Direct May Not Be the Best: An Incremental Evolution View of Pose Generation

AAAI 2024 · 3D vision › Human pose estimation · approach: Learned camera pose estimation · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes an incremental evolution pose generation framework that synthesizes target poses through a sequence of small intermediate pose steps, guided by global skeleton/semantic evolution constraints and a triple-path knowledge fusion network.

Problem

Human pose transformation from a source image to a target pose is hard because direct one-to-one mapping must model large non-linear content variation, causing distortion and blur.

Approach

Replaces direct one-to-one pose transfer with a gentle incremental evolution course of small pose steps; global RNN-based skeleton/semantic evolution constraints and incremental evolution blocks guide each step, and a triple-path knowledge fusion (TPKF) block fuses source, incremental constraints and target information.

Key results

• On Turning-Round, Evolution View-L reaches SSIM 0.953, PSNR 24.265, FID 51.384, LPIPS 0.053, beating the best baseline SPGNet (0.949, 23.949, 56.853, 0.056) and PISE.
• On Fashion, Evolution View-L achieves SSIM 0.930, PSNR 25.332, FID 12.720, LPIPS 0.048, the best among the compared methods.
• In a human perceptual study on Fashion, the method reaches R2G 42.4% and G2R 20.8%, higher than PISE, CASD and BiGraphGAN.

Measured effects reported in the paper

Surprising findings

• Stacking more IE blocks (six or nine) degrades performance relative to three, which the authors attribute to loss of critical details from overly deep convolution fusion.
• The method has high computational cost in MACs (131-464 G) compared with several baselines, though its parameter count is moderate.

Limitations

• Relatively high MACs due to the incremental multi-step generation.
• Ablation shows that adding more IE blocks hurts performance, indicating sensitivity to depth.
• Loss weights are fixed, and the authors report sensitivity to weight fluctuations (up to 33% FID change).

Experimental setup

Datasets: Fashion Tai-Chi Turning-Round

Benchmarks: Turning-Round Fashion Tai-Chi

Metrics: SSIM PSNR FID LPIPS #Param MACs

Baselines: PATN XingGAN ADGAN PINet PISE SPGNet CASD DPTN BiGraphGAN MAGPT PIDM

Builds on: GAN Transformer AdaIN OpenPose

Closest papers by meaning

• Disentangled Pose and Appearance Guidance for Multi-Pose Generation (CVPR 2025 · similarity 0.86)
• Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.84)
• RecDreamer: Consistent Text-to-3D Generation via Uniform Score Distillation (ICLR 2025 · similarity 0.84)
• FlexPose: Pose Distribution Adaptation with Limited Guidance (AAAI 2025 · similarity 0.83)
• DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images (NeurIPS 2026 · similarity 0.83)
• Leveraging Global Stereo Consistency for Category-Level Shape and 6D Pose Estimation from Stereo Images (CVPR 2025 · similarity 0.83)