Why can't you just trust the numbers?
Quantitative metrics like PSNR (peak signal-to-noise ratio) and SSIM (structural similarity) measure pixel-level differences, but they don't tell you if the reconstruction looks natural or preserves identity. In [2], the authors report improvements in PSNR, SSIM, LPIPS (learned perceptual image patch similarity), and FID (Fréchet inception distance), but they also emphasize qualitative assessments for 'clothing details, identity consistency, and pose accuracy'—things numbers alone can't capture. So, a high PSNR might still look 'off' to a human eye if the clothing wrinkles are wrong or the face doesn't match.
Similarly, [1] highlights that their method captures 'subtle geometry of clothing' using image-based priors like surface normals and optical flow, but the paper's evaluation includes qualitative comparisons because these details are hard to quantify. The takeaway: use metrics as a first pass, but always pair them with human judgment.
What about motion and time?
For 4D reconstruction, temporal consistency—how smoothly the mesh or appearance changes over frames—is crucial. [3] specifically addresses 'fine-grained dynamic details, particularly in clothing wrinkles and hair motion,' which are often lost in static metrics. They propose a residual deformation field to capture these details, but evaluating them requires watching the sequence over time, not just looking at single frames.
One practical approach is to render novel views from the reconstruction and ask viewers to rate temporal flicker or jitter. [5] also notes that their method handles 'large deformations' and 'shape completion in occluded regions,' which are subjective aspects that need visual inspection. So, when evaluating, include time-lapse videos or interactive viewing to assess motion quality.
How do you actually run a subjective evaluation?
A practical protocol combines automated metrics with structured user studies. Start with standard metrics: PSNR, SSIM, LPIPS for image quality, and FID for distribution similarity, as used in [2]. Then, recruit a small group of raters (e.g., 10–20) to view side-by-side comparisons of reconstructions from different methods, rating on criteria like 'identity preservation,' 'clothing detail,' and 'temporal stability.' This mirrors what [1] and [3] implicitly do in their qualitative comparisons.
Also, consider task-specific evaluation: if the goal is telepresence, as in [2], ask users to judge 'presence' or 'realism' in a conversational scenario. If it's for animation, as in [4], test whether the reconstructed mesh can be re-posed without artifacts. The key is to align the evaluation with the intended use case.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2023 to 2025, 3 from 2024 or later, collectively cited 179 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 42 papers retrieved from a database of over 500 million.
Sources used in this answer
DressRecon: Freeform 4D Human Reconstruction from Monocular Video
DressRecon combines human priors with video-specific deformation to capture loose clothing and object interactions, and reports higher-fidelity 3D reconstructions than prior art on challenging datasets, but does not provide quantitative metrics in the abstract.
Monocular 3-D Human Reconstruction Using Gaussian Splatting Based On Multiview Priors
The thesis presents a real-time monocular reconstruction pipeline using Gaussian splatting and multiview priors, achieving improvements in PSNR, SSIM, LPIPS, and FID over state-of-the-art, with qualitative assessments showing better clothing detail and identity consistency.
Animatable NeRF Dynamic Detail Enhancement Based on Residual Deformation Field with Progressive Training
The paper introduces a residual deformation field and progressive training to enhance dynamic details like clothing wrinkles and hair motion in animatable NeRF, but does not report quantitative metrics in the abstract.
Humans in 4D: Reconstructing and Tracking Humans with Transformers
4DHumans uses a transformer-based HMR 2.0 for mesh recovery and 3D tracking, achieving state-of-the-art tracking results from monocular video, but the abstract focuses on tracking accuracy rather than subjective quality.
Unbiased 4D: Monocular 4D Reconstruction with a Neural Deformation Model
Ub4D proposes an unbiased neural deformation model for monocular 4D reconstruction, demonstrating improvements in surface reconstruction accuracy and robustness to large deformations, but does not include subjective quality metrics in the abstract.