kfkas/streaming-gebd-causal
Audited Kinetics-GEBD Causal Metadata This metadata-only release converts the publicly released Kinetics-GEBD annotations into an auditable 24 FPS causal training representation. It does not redistribute Kinetics or YouTube video bytes. Splits Hub split Official source file Records Meaning train k400_train_raw_annotation.pkl 18,808 Public GEBD training annotations validation k400_val_raw_annotation.pkl 18,815 Public Kinetics-GEBD validation… See the full description on the dataset page: https://huggingface.co/datasets/kfkas/streaming-gebd-causal.
Audited Kinetics-GEBD Causal Metadata
This metadata-only release converts the publicly released Kinetics-GEBD annotations into an auditable 24 FPS causal training representation. It does not redistribute Kinetics or YouTube video bytes.
Splits
The ICCV 2021 benchmark describes a separate challenge test subset sampled from Kinetics-400 train. Its labels are withheld and are not included here. The ICCV 2025 On-GEBD paper explicitly evaluates Kinetics-GEBD on the public validation split because test annotations are unavailable.
Annotation processing
The release reproduces the official `prepare_k400_release.ipynb` behavior, including its input-order and list-mutation quirks:
- discard boundaries in the first or last 0.3 seconds;
- merge overlapping gradual-shot ranges;
- suppress timestamp boundaries around gradual-shot ranges;
- suppress nearby same-type timestamps and prefer shots over events;
- convert cleaned timestamps to source-frame indices with
floor(timestamp * source_fps), capped atnum_frames - 1.
The implementation was differentially checked against the official notebook over all 185,518 released rater annotations with zero timestamp-sequence differences. This is intentionally the code behavior: some effective suppression distances are 0.6 seconds even though the paper summarizes a 0.1-second merge rule.
Training uses the first maximum of f1_consis, matching numpy.argmax in the official loader. Evaluation must compare predictions independently with every retained rater and select the best rater-specific F1. Records with f1_consis_avg < 0.3 remain present but have evaluation_eligible=false.
Causal 24 FPS labels
The benchmark does not publish a canonical 24 FPS binary-label file. This release derives one for causal training:
sample_count = ceil(duration_seconds * 24);- sample
irepresents timestampi / 24; - a sample is positive when it falls inclusively within 0.15 seconds of a cleaned boundary from the highest-consistency training rater.
boundary_source_frame_indices and training_boundary_source_frame_indices preserve official source-FPS frame coordinates. training_boundary_sample_indices and binary_labels are the derived 24 FPS projection. These fields must not be conflated.
Evaluation
The official metric is relative temporal distance, not the 0.15-second training window. The paper reports F1 at thresholds 0.05 through 0.50; the challenge ranking metric is F1@0.05. Matching is one-to-one and performed separately for each rater, retaining the rater yielding the best F1.
Videos
The local acquisition audit resolved and ffprobe-validated 37,608 of 37,623 referenced clips:
The remaining 15 references were absent from all inspected public mirrors, the official CVDF Kinetics-400 archives, and accessible original-video paths. See `VIDEO_ACQUISITION.md` and manifests/ for exact availability and missing-reference records.
uv run scripts/build_official_on_gebd_dataset.py
uv run scripts/acquire_official_gebd_videos.py --workers 4Rights
The GEBD annotation release is CC BY-NC 4.0 and the official repository code is MIT-licensed. Neither license grants redistribution rights for the underlying Kinetics/YouTube videos. A gated or private Hub repository does not change those rights, so this release contains metadata and manifests only.