🧑‍🤝‍🧑 Crowd-3DGS Dataset

A sequential-frame dataset of a real-world crowd protest scene, fully prepared for 3D Gaussian Splatting (3DGS) reconstruction — including high- and low-resolution image sets, COLMAP sparse reconstruction, stereo depth maps, and ready-to-run COLMAP scripts.

---

Dataset Description

Crowd-3DGS is a single-scene video-derived dataset capturing a crowd protest, intended for 3D scene reconstruction and novel view synthesis using 3D Gaussian Splatting (3DGS). The dataset ships as a complete, pipeline-ready package: sequential JPG frames extracted from video, camera calibration via COLMAP Structure-from-Motion (SfM), stereo dense reconstruction outputs, and helper shell scripts to re-run COLMAP from scratch.

Crowd and protest scenes present unique challenges for 3DGS: heavy human occlusion, transient moving subjects, non-rigid motion, and complex lighting. This dataset provides a realistic real-world benchmark for researchers tackling robust 3DGS under these conditions.

• Curated by: siyah1
• Funded by:
• Language(s): N/A (visual/3D dataset)
• License: CC-BY 4.0
• Total size: ~392 MB

Related Links

Resource	Link
🤗 Dataset Repository	siyah1/crowd-3dgs
📄 Paper
💻 Code / Training Script
🌐 Project Page

---

Background: What is 3D Gaussian Splatting?

3D Gaussian Splatting (3DGS) (Kerbl et al., 2023) is an explicit 3D scene representation that models a scene as a collection of anisotropic 3D Gaussian primitives, each defined by:

• Position (mean): 3D location in space
• Covariance (shape/orientation): size and orientation of the Gaussian
• Opacity: transparency level
• Spherical Harmonics (SH) coefficients: view-dependent colour and appearance

3DGS achieves real-time rendering at high visual fidelity and is widely used in VR, AR, robotics, and digital heritage. However, crowd scenes — with heavy occlusion, dynamic humans, and transient elements — remain an open challenge for the method.

---

Uses

Direct Use

• 3D reconstruction of a real-world crowd/protest scene using 3DGS or NeRF-based methods
• Novel view synthesis benchmarking in challenging, densely populated environments
• Research into transient object handling, occlusion-robust 3DGS, and dynamic scene reconstruction
• Prototyping end-to-end COLMAP-to-3DGS pipelines using the included scripts and pre-processed data

Out-of-Scope Use

This dataset is not intended for:

• Facial recognition, individual identity tracking, or surveillance
• Training crowd-counting or person-detection models (see CrowdHuman, UCF-QNRF for those)
• Commercial deployment without appropriate licensing clearance
• Any use that violates the privacy or dignity of individuals depicted

---

Dataset Structure

File Tree

crowd-3dgs/                          (~392 MB total)
│
├── images/                          (~202 MB) ← High-resolution frames
│   ├── frame_0001.jpg               (~675 kB)
│   ├── frame_0002.jpg               (~664 kB)
│   ├── frame_0003.jpg               (~666 kB)
│   └── ...                          (sequential video frames)
│
├── input/                           (~95.1 MB) ← Low-resolution frames
│   ├── frame_0001.jpg               (~324 kB)
│   ├── frame_0002.jpg               (~314 kB)
│   ├── frame_0003.jpg               (~316 kB)
│   └── ...                          (same sequence, downscaled)
│
├── sparse/                          ← COLMAP SfM output
│   └── 0/
│       ├── cameras.bin              (camera intrinsics)
│       ├── images.bin               (per-frame pose / extrinsics)
│       └── points3D.bin             (sparse 3D point cloud)
│
├── stereo/                          (~12.9 kB) ← COLMAP stereo config files
│   ├── fusion.cfg                   (4.5 kB — stereo fusion settings)
│   └── patch-match.cfg              (8.4 kB — patch-match stereo settings)
│
├── run-colmap-geometric.sh          ← Script: geometric stereo fusion
├── run-colmap-photometric.sh        ← Script: photometric stereo fusion
├── .gitattributes
└── README.md

Folder Descriptions

Folder / File	Size	Description
images/	~202 MB	Full-resolution sequential JPG frames from source video. Primary input to 3DGS.
input/	~95.1 MB	Downscaled version of the same frames. Useful for faster COLMAP or low-VRAM training.
sparse/	—	COLMAP SfM output: camera intrinsics, per-frame poses, sparse 3D point cloud for 3DGS initialisation.
stereo/	~12.9 kB	COLMAP stereo configuration files: fusion.cfg (fusion settings) and patch-match.cfg (patch-match stereo settings). Run the included .sh scripts to generate depth/normal maps.
run-colmap-geometric.sh	—	Re-run COLMAP dense stereo with geometric consistency filtering.
run-colmap-photometric.sh	—	Re-run COLMAP dense stereo with photometric consistency filtering.

Image Statistics

Property	images/ (hi-res)	input/ (lo-res)
Format	JPEG	JPEG
Avg. file size	~660 kB	~315 kB
Naming convention	frame_XXXX.jpg	frame_XXXX.jpg
Scene	Crowd protest	Crowd protest

Frame count:

---

How to Use

Quick Start with 3DGS

The sparse/ folder already contains pre-computed COLMAP output, so you can jump straight to training:

# 1. Clone the dataset
git clone https://huggingface.co/datasets/siyah1/crowd-3dgs

# 2. Train with the official 3DGS repo (https://github.com/graphdeco-inria/gaussian-splatting)
python train.py -s crowd-3dgs/

Re-running COLMAP Dense Stereo (Optional)

# Geometric stereo (recommended for cleaner reconstructions)
bash run-colmap-geometric.sh

# Photometric stereo (alternative)
bash run-colmap-photometric.sh

Using input/ for Fast Iteration

The downscaled input/ frames (~315 kB vs ~660 kB) are useful for:

• Rapid prototyping without loading the full 202 MB image set
• COLMAP runs on machines with limited RAM
• Initial hyperparameter sweeps before full-resolution training

---

Dataset Creation

Curation Rationale

Standard 3DGS benchmarks (Tanks & Temples, Mip-NeRF 360, Deep Blending) are dominated by static or near-static scenes with sparse human presence. Crowd-3DGS targets the gap for densely-populated, real-world protest/event scenarios — where occlusion, transient people, and motion make reconstruction significantly harder and more practically relevant.

Source Data

Data Collection and Processing

• Scene type: Outdoor crowd protest
• Capture method: [e.g., handheld smartphone / DSLR video — More Information Needed]
• Frame extraction: Sequential frames extracted from video (frame_0001.jpg, frame_0002.jpg, …)
• Camera calibration: COLMAP SfM run on input/ frames; camera poses stored in sparse/0/
• Dense reconstruction: COLMAP patch-match stereo configured via stereo/patch-match.cfg and stereo/fusion.cfg; run the included .sh scripts to generate depth/normal maps
• Multi-resolution: Full-res in images/, downscaled copies in input/

Who are the source data producers?

Video captured by [More Information Needed]. All individuals appear in a publicly filmed protest scene.

Personal and Sensitive Information

This dataset contains images of people participating in a public protest:

• Anonymization applied: [Yes / No — More Information Needed]
• EXIF metadata: Verify GPS/timestamp data is stripped before publishing
• This dataset is not suitable for individual identification, political tracking, or surveillance
• Protest participation is a form of political expression — treat subjects' dignity accordingly

⚠️ If face anonymization has not been applied, add appropriate usage restrictions and warnings.

---

Bias, Risks, and Limitations

Technical Limitations

Issue	Impact on 3DGS
Transient moving people	Inconsistency across frames → floaters, ghosting, blurring in Gaussians
Video-derived viewpoints	Camera path is sequential, not a multi-rig; viewpoint diversity is limited
Motion blur	Fast crowd movement may degrade per-frame sharpness
Single scene	Results may not generalise to other crowd types, densities, or locations
Stereo depth noise	Dense depth maps from patch-match are noisy on crowd/human regions

Sociotechnical Risks

• Individuals are participants in a public protest — a politically sensitive context
• Use for surveillance, identity recognition, or political profiling is explicitly prohibited
• Geographic and demographic biases exist based on the specific protest location and demographics

Recommendations

• Apply transient masking during 3DGS training (semantic segmentation of humans, or RobustNeRF-style robust loss)
• Compute PSNR / SSIM / LPIPS on static background regions, masking dynamic foreground humans
• Do not use this dataset or derived models for crowd monitoring, surveillance, or profiling

---

Evaluation

Recommended metrics for benchmarking 3DGS on this dataset:

Metric	Description	Direction
PSNR	Peak Signal-to-Noise Ratio	↑ Higher is better
SSIM	Structural Similarity Index	↑ Higher is better
LPIPS	Learned Perceptual Image Patch Similarity	↓ Lower is better
FPS	Rendering speed	↑ Higher is better

---

Citation

BibTeX:

@dataset{crowd3dgs2026,
  author    = {siyah1},
  title     = {Crowd-3DGS: A Sequential Frame Dataset of a Crowd Protest Scene for 3D Gaussian Splatting},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/siyah1/crowd-3dgs}
}

APA:

siyah1. (2026). Crowd-3DGS: A Sequential Frame Dataset of a Crowd Protest Scene for 3D Gaussian Splatting [Dataset]. Hugging Face. https://huggingface.co/datasets/siyah1/crowd-3dgs

---

Glossary

Term	Definition
3DGS	3D Gaussian Splatting — real-time novel view synthesis using anisotropic 3D Gaussian primitives
NVS	Novel View Synthesis — rendering a scene from camera positions not seen during training
SfM	Structure-from-Motion — recovering 3D camera poses from 2D images
COLMAP	Open-source SfM + MVS pipeline; standard preprocessing tool for 3DGS
Sparse reconstruction	COLMAP's SfM output: sparse 3D points + per-frame camera poses
Dense / stereo reconstruction	Patch-match stereo producing per-pixel depth and normal maps
Geometric stereo	COLMAP mode enforcing cross-frame geometric consistency
Photometric stereo	COLMAP mode enforcing cross-frame photometric (appearance) consistency
Transient objects	Dynamic elements (moving people) that appear inconsistently across frames
Spherical Harmonics (SH)	Basis functions encoding view-dependent colour in 3DGS

---

Dataset Card Contact

siyah1 on Hugging Face

---

Fields marked "[More Information Needed]" should be completed by the dataset author. Card follows the HuggingFace dataset card spec.