SoccerNet/SN-Features

SoccerNet Features

Pre-extracted per-game features for the SoccerNet benchmark, structured as <league>/<season>/<game>/<file>, one file per game half (1_.../2_...).

This main branch holds no data — each feature type lives on its own branch so you only download what you need:

Branch	Files	Description
baidu-soccer-embeddings	{1,2}_baidu_soccer_embeddings.npy	Frame embeddings from baidu-research/vidpress-sports, used by the Action Spotting and Dense Video Captioning 2023 challenges
resnet-tf2	{1,2}_ResNET_TF2.npy	ResNET features @2fps, extracted with TF2 (SoccerNetv2-DevKit)
resnet-tf2-pca512	{1,2}_ResNET_TF2_PCA512.npy	Same as above, dimensionality-reduced to 512 with PCA
player-boundingbox-maskrcnn	{1,2}_player_boundingbox_maskrcnn.json	Player bounding boxes @2fps, extracted with MaskRCNN
field-calib-ccbv	{1,2}_field_calib_ccbv.json	Field camera calibration @2fps, extracted with CCBV

Download

Using the SoccerNet pip package (recommended — matches the local folder layout used by the rest of the SoccerNet.Downloader API):

from SoccerNet.Downloader import SoccerNetDownloader
d = SoccerNetDownloader(LocalDirectory="path/to/soccernet")
d.downloadDataTask(task="spotting-2023", split=["train", "valid", "test", "challenge"])
d.downloadDataTask(task="caption-2023", split=["train", "valid", "test", "challenge"])

Directly with huggingface_hub, picking a branch and (optionally) a subset of games via allow_patterns:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="SoccerNet/SN-Features",
    repo_type="dataset",
    revision="resnet-tf2-pca512",   # one of the branches listed above
    local_dir="path/to/soccernet",
)

Corresponding labels (Labels-v2.json, Labels-caption.json) are in SoccerNet/SN-Labels. Held-out test/challenge ground truth is in the private SoccerNet/SN-GroundTruth.