FineDiving trimmed RGB frames

Personal mirror of the FineDiving trimmed-frame material obtained from OpenDataLab/FineDiving. The original dataset and annotations are by the FineDiving authors. This mirror does not claim authorship of the source data.

Contents

• 3,000 trimmed clips across 52 diving action classes.
• 312,256 JPEG frames, stored in one ZIP (4,736,265,072 bytes).
• Four source annotation files, also available separately under annotations/.
• Historical official split lists: 2,251 training clips and 749 test clips.
• Source provenance in metadata/, plus MANIFEST.json and SHA256SUMS.

The ZIP preserves the source frame paths:

FINADiving_MTL_256s/<group>/<clip>/<frame>.jpg
transfer_annotations/<annotation files and source notes>

All 3,000 clips are included. The archive has not been filtered to the test set. The four annotations embedded in the ZIP are duplicated as small standalone files for convenient access. Official train/test lists are provided separately in annotations/; they are not embedded in the ZIP.

Untrimmed full-length videos are outside the scope of this mirror. No reliable physical frame rate is asserted by this package. Frame sampling and timing must be configured explicitly by downstream users.

Official split provenance

The split lists are the unchanged files from author repository commit 273bcdb7a3e438b15356b2f9e459652daaf712ce. They contain 2,251 and 749 unique IDs, have no overlap, and jointly cover all 3,000 annotated clips. The original paper describes the ratio approximately as 75% training and 25% test.

The official protocol does not define a Base/Novel class partition. Such a partition is a separate downstream evaluation choice; no custom Base/Novel protocol is included in this mirror.

Download and extract

For this private repository, authenticate with an account that has access first:

hf auth login
hf download ZWJ666/FineDiving --repo-type dataset --local-dir FineDiving
cd FineDiving
sha256sum -c SHA256SUMS
unzip FineDiving_trimmed_frames_with_annotations.zip

On macOS, use shasum -a 256 -c SHA256SUMS for checksum verification. Select clips using annotations/train_split.pkl or annotations/test_split.pkl before training or evaluation. Each ID is a (group, clip) pair corresponding to FINADiving_MTL_256s/<group>/<clip>/.

Source and citation

Source archive member: FineDiving/Released_FineDiving_Dataset/Trimmed_Video_Frames/FineDiving_Trimmed_VideoFrames.zip. Only the trimmed-frame ZIP and accompanying small annotations were downloaded from the larger OpenDataLab archive. The full parent archive was not downloaded or hashed; its expected checksum is retained only as source metadata. The local trimmed ZIP was CRC checked, and all transferred files have SHA-256 checksums.

Please cite the original paper, FineDiving: A Fine-grained Dataset for Procedure-aware Action Quality Assessment, and consult the original authors' dataset terms for use and redistribution.