SoccerNet Novel View Synthesis - SN-NVS

Novel view synthesis has transformed 3D scene reconstruction, enabling the rendering of photorealistic perspectives from sparse observations and democratizing immersive content creation. Soccer broadcasts can offer a interesting application for novel view synthesis (NVS), as they provide multiple camera angles of dynamic scenes with complex interactions. However, the gap between broadcast point of view and novel viewpoints using the point of view of a player or a referee remains significant, making it challenging to generate high-quality novel views in soccer scenarios.

Dataset

The dataset contains 5 scenes generated with the Blender game engine, featuring different soccer actions. Each scene is divided in 2 different sets: train and challenge. The train set includes broacasted images and camera parameters using COLMAP with known poses. For the challenge set, only the poses of the novel views are provided, but not the ground truth images. However, for the sake of compatibility with existing NVS methods, we provide dummy images (black images) in the challenge set.

The dataset follow the structure below:

SoccerNet-NVS/
├── scene-1/
│   ├── images/
│   ├── images_2/
│   ├── images_4/
│   ├── images_8/
│   └── sparse/
│       └── 0/
│           ├── cameras.bin
│           ├── images.bin
│           ├── points3D.bin
│           └── points3D.ply
├── scene-2/

Note that the original images/ have a 4K resolution (4096x2160), and the other folders contain downsampled versions. Participants are free to use any of these resolutions for training, but the evaluation will be performed at a resolution of 2048x1080 (i.e., images_2/).

The dataset is available at our Hugging Face repository: https://huggingface.co/datasets/SoccerNet/SN-NVS-2026/tree/main

Evaluation

We evaluate the reconstruction quality using the PSNR following standard practices. However, we also encourage the participants to keep track of other metrics such as SSIM and LPIPS, as they provide complementary information about the reconstruction quality of the generated images. We will evaluate the images using a resolution of 2048x1080 (equivalent to images_2).

Baselines

We evaluated 2 baseline methods: 3D Gaussian Splatting 3DGS and Triangle Splatting TS.
The results on the challenge set are the the following:

Method	PSNR ↑	SSIM ↑	LPIPS ↓
3DGS	26.74	0.75	0.41
TS	26.43	0.757	0.359

Training the Baselines

After you have downloaded the dataset and installed the chosen method, you can train the baseline methods using the following commands.

Installing the repository

Here are the basic steps to install the 3DGS repository (same for TS) from this repository, as submodules:

git clone https://github.com/SoccerNet/sn-nvs --recursive
cd submodules/gaussian-splatting

conda env create --file environment.yml
conda activate gaussian_splatting

Training

cd submodules/gaussian-splatting
conda activate gaussian_splatting
python train.py -s <path_to_scene> -i images_2 -r 1 -m <path_to_save_model> --eval --extra_flags

with --extra_flags being related to the specific method.

Rendering

cd submodules/gaussian-splatting
conda activate gaussian_splatting
python render.py -s <path_to_scene> -m <path_to_trained_model> --eval --skip_train

Validation set metrics

cd submodules/gaussian-splatting
conda activate gaussian_splatting
python metrics.py -m <path_to_trained_model>

Codabench

For the challenge submission, we use Codabench as the submission platform. You can find the challenge page here: https://www.codabench.org/competitions/11339/.

For the challenge submission, participants should use the provided render_challenge.py script (equivalent to render.py from 3DGS). To use it, you need to specify the model path you used during training, but you need to specify the challenge scene path. An example can be found below, using the 3DGS method:

conda activate gaussian_splatting
python render_challenge.py 3dgs -s <path_to_challenge_scene> -m <path_to_trained_model> -i images_2 -r 1

Depending on the method you use, you may need to change the render function. You can also directly copy and modify the render_challenge.py script to your own repository to fit your method.

The participants need to submit their results in a zip file with the following structure:

├── scene-1-challenge/
│   ├── renders/
│   │   ├── 000000.png
│   │   ├── 000001.png
│   │   └── ...
├── scene-2-challenge/
│   ├── renders/
│   │   ├── 000000.png
│   │   ├── 000001.png
│   │   └── ...

The baseline results can be found in the Hugging Face repository: https://huggingface.co/datasets/SoccerNet/SN-NVS-2026/tree/main

Additional Information

If your method is based on the 3D Gaussian Splatting codebase, do not forget to use both -i images_2 and -r 1 flags to ensure the correct resolution and to disable the automatic rescaling during training.