💻 computer science

From Far and Near: Perceptual Evaluation of Crowd Representations Across Levels of Detail

This paper investigates how users perceive the visual quality of various crowd representations—including meshes, impostors, NeRFs, and 3D Gaussians—across different levels of detail and viewing distances to provide insights for designing perceptually optimized rendering strategies.

Original authors: Xiaohan Sun, Carol O'Sullivan

Published 2026-03-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaohan Sun, Carol O'Sullivan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are directing a massive movie scene with a crowd of thousands of people. You want the scene to look incredibly realistic, but your computer (or the gaming console) can't handle rendering every single detail for every single person at once. If it tried, the movie would freeze, or the game would run at a snail's pace.

To solve this, directors use a trick called Levels of Detail (LoD). Think of it like a camera zooming in and out:

Close up: When a character is right in front of the camera, you see every wrinkle, hair strand, and fabric fold (High Detail).
Far away: When a character is 100 meters away, you don't need those details. A blurry blob or a flat picture of the person is enough for the human eye to believe it's a real person (Low Detail).

This paper is a scientific "taste test" to figure out which method of faking a crowd works best at which distance.

The Four "Actors" in the Study

The researchers compared four different ways to render these crowd members, like four different actors auditioning for the same role:

The Geometric Mesh (The Classic Actor):
- What it is: A 3D model made of thousands of tiny triangles (like a digital wireframe covered in skin).
- Pros: It looks amazing up close and moves naturally.
- Cons: It's heavy. It takes a lot of computer power to render, especially if you have thousands of them.
The Impostor (The Flat Poster):
- What it is: Instead of a 3D model, it's just a flat 2D picture (a sprite) that always faces the camera. It's like a cardboard cutout that spins to face you.
- Pros: Super fast and light on the computer.
- Cons: If you walk around it, it looks weird because it has no depth. It's a flat image pretending to be 3D.
The NeRF (The Magic Paintbrush):
- What it is: A new AI technique (Neural Radiance Fields) that "learns" what the person looks like from every angle and recreates it using math.
- Pros: Can look incredibly realistic and smooth.
- Cons: It's computationally heavy to "teach" the AI, and rendering it can be tricky.
The 3D Gaussian (The Cloud of Dots):
- What it is: Imagine the person is made of millions of tiny, fuzzy, glowing balloons (Gaussians) floating in space.
- Pros: Very fast to render and looks great from many angles.
- Cons: It can look a bit "sparkly" or blurry if you zoom in too close.

The Experiment: The "Blind Taste Test"

The researchers set up a controlled experiment. They showed people a perfect, high-quality 3D model (the "Gold Standard") and then asked them to guess which of the four "fake" versions looked most like the real thing.

They tested this under two conditions:

Distance: Is the person standing right next to you, or are they far away?
Motion: Is the person standing still, or are they walking?

The Big Findings (The "So What?")

Here is the simple breakdown of what they discovered:

1. Distance is King

Close Up: When you are close to a character, the Geometric Mesh (the classic 3D model) is the undisputed champion. It's the only one that doesn't look "fake."
Far Away: As the character gets further away, the 3D Gaussians and Impostors become almost indistinguishable from the real thing. Your brain can't tell the difference anymore.

2. Motion Doesn't Matter (As Much as You Think)
Surprisingly, whether the character was walking or standing still didn't change the results much. If a fake version looks good when the person is still, it usually looks good when they are walking, too.

3. The "Sweet Spot" for Each

Use Impostors for the background: If you need to render 1,000 people in the distance, use the flat "poster" method. It saves massive amounts of computer power and looks fine from far away.
Use 3D Gaussians for the middle ground: They are the new "Goldilocks" option. They look great, are fast to render, and are becoming a favorite for modern games.
Use Meshes for the foreground: If a character is the main focus or close to the player, stick with the heavy, detailed 3D model.

4. The Trade-off

Meshes are heavy but need no "training."
Impostors are tiny and fast but require pre-made pictures.
NeRFs and Gaussians are the new kids on the block. They offer a great balance of quality and speed, but they require a bit of "training time" (like teaching an AI) before you can use them.

The Takeaway for the Future

This paper gives game developers and VR creators a "recipe book." Instead of guessing which method to use, they can now say:

"Okay, for characters closer than 5 meters, we'll use the heavy 3D models. For characters between 5 and 20 meters, we'll switch to the fast 3D Gaussians. And for everyone else in the background? We'll use the tiny, flat posters."

By mixing and matching these methods based on how far away the player is, we can create massive, realistic crowds in video games and virtual reality without making the computer overheat. It's about getting the most "bang for your buck" in visual quality.

1. Problem Statement

Crowd rendering requires balancing visual realism with computational performance, particularly in real-time applications like games and VR. Traditionally, Levels of Detail (LoD) strategies have relied on geometric meshes (simplified via decimation) and image-based impostors (2D sprites). However, the emergence of neural rendering techniques, specifically Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), introduces new representation options.

The core research question is: How do traditional LoD approaches (Mesh, Impostor) compare with modern neural representations (NeRF, 3DGS) in terms of perceptual quality across varying viewing distances and levels of detail? Specifically, at what point do neural representations become perceptually indistinguishable from high-fidelity ground-truth meshes, and how do motion and distance influence these thresholds?

2. Methodology

A. Dataset and Representations

The authors constructed a unified dataset using a single Mixamo female character walking for 60 frames, captured from 60 virtual cameras on a hemispherical rig (3,600 images total). They instantiated four distinct crowd representations from this shared data:

Geometric Mesh (M): The baseline high-fidelity representation. Four LoDs (L0–L3) were generated using Blender's Decimate modifier, reducing face counts from ~27k (L0) to ~5.8k (L3).
Image-based Impostor (I): Precomputed 2D sprite sheets (6×10 tiles) per LoD. L0 resolution was 1080×1080px, halving at each subsequent level. To prevent "breathing" artifacts, a union alpha bounding box was used to stabilize frames before packing.
Neural Radiance Fields (NeRF): Implemented using Instant-NGP. LoDs were controlled by varying the hash table capacity and MLP width. Independent NeRFs were trained for each of the 60 frames.
3D Gaussian Splatting (3DGS): Implemented using the standard 3DGS pipeline. LoDs were controlled by capping the Gaussian count ( $N$ ) from 120k (L0) down to 1.9k (L3).

B. Experimental Design (User Study)

A controlled user study was conducted with 24 participants.

Stimuli: Participants viewed four representations (M, I, N, G) side-by-side at four LoDs (100%, 50%, 25%, 12.5%) and five viewing distances (100% to 20% of screen pixels).
Conditions: Two modes were tested: Video (animated sequence) and Image (static frame).
Task: Participants selected which of the four stimuli most closely matched a high-resolution mesh reference.
Analysis: Data was analyzed using OLS ANOVA and binomial GLM (logit) with subject fixed effects to determine statistical significance of factors: Representation, Distance, LoD, and Mode (Motion).

C. Quantitative Evaluation

Beyond user perception, the authors performed objective image fidelity comparisons against the Mesh baseline using PSNR, SSIM, and LPIPS. They also measured on-disk asset sizes and offline training times (for learning-based methods) to evaluate deployment efficiency.

3. Key Contributions

Perceptual Thresholds for LoD Placement: The study provides empirical data on when different representations are perceptually indistinguishable from high-quality meshes. It establishes that 3D Gaussians become increasingly similar to meshes as detail decreases or viewing distance increases, whereas Impostors maintain structural fidelity at a distance but lack 3D parallax.
Perception-Driven LoD Pipeline: The authors introduce a representation-aware rendering pipeline and toolkit that integrates Mesh, Impostor, NeRF, and 3DGS. They provide practical guidelines for generating customized LoD assets for each type.
Comprehensive Comparative Analysis: This is one of the first works to systematically compare traditional and neural crowd representations across a matrix of LoDs and viewing distances, including both static and dynamic scenarios.

4. Results

Perceptual Findings

Main Effects: Perceived similarity is driven primarily by Representation, Viewing Distance, and LoD.
Motion Independence: Surprisingly, the presence or absence of motion (Mode) had no significant effect on perceptual judgments. Participants' ability to distinguish representations remained consistent between static images and video.
Representation Performance:
- Mesh: Preferred at high detail (L0) and near distances.
- 3D Gaussians: Become perceptually indistinguishable from meshes as resolution drops or distance increases. They offer a strong "mid-to-high" LoD alternative.
- Impostors: Highly competitive at far distances and low LoDs due to preserved silhouettes, though they suffer from a lack of 3D parallax.
- NeRF: Degrades smoothly in PSNR but maintains high structural similarity (SSIM) even at lower LoDs.

Quantitative & Efficiency Findings

Fidelity Metrics: At L0, both 3DGS and NeRF closely match the Mesh (PSNR > 36 dB, SSIM > 0.98). As LoD decreases, NeRF maintains high SSIM, while 3DGS shows a larger PSNR drop but remains competitive on structural metrics. Impostors show lower PSNR but very low LPIPS (high perceptual similarity) at matched views.
Memory & Training:
- Impostors: Most memory-efficient (KB scale) and require zero training time.
- 3DGS: Highly compressible (MB scale at L3) with training time decreasing as complexity drops.
- Mesh: Moderate memory savings with simplification; no training cost.
- NeRF: Largest deployment footprint and consistent training costs per frame.

5. Significance and Future Work

Practical Guidelines: The paper offers actionable rules for developers:
- Use Impostors for far-field, low-detail rendering to maximize batching and minimize memory.
- Use 3D Gaussians for mid-to-high LoD ranges where real-time rendering and visual fidelity are balanced.
- Use Geometric Meshes for close-up, high-detail views.
Impact on VR/Immersive Tech: These findings allow for perceptually optimized LoD strategies that maintain frame rates and user comfort in crowd-populated virtual environments without sacrificing visual quality.
Future Directions: The authors suggest extending this work to diverse character appearances, complex lighting, physical interactions (clothing dynamics), and end-to-end runtime performance benchmarks to fully quantify the trade-offs in production environments.

In conclusion, this paper bridges the gap between traditional graphics and neural rendering by providing a perceptual framework for selecting the optimal crowd representation based on viewing conditions, proving that neural methods (specifically 3DGS) are viable, high-fidelity alternatives to traditional meshes in specific LoD regimes.