Access is restricted to individually approved researchers. Applications are reviewed manually. Read the Data Use Agreement before submitting. Review time varies.

By submitting this application, you accept the ProAct Data Use Agreement and share your application information with the dataset maintainers for access review.

ProAct Self-Collected Research Release

This ProAct release contains privacy-masked self-collected recordings for non-commercial scientific research, with individual manual access approval. Read the publicly accessible Data Use Agreement and application instructions before applying. Both documents are readable before Hugging Face approval.

Item	Included
Masked videos	495 at 3840×2160, with no audio track
Masked training keyframes	18,747 JPEGs at 768×432
Takes / annotated tasks	165 / 15
Annotation records / step segments	569 / 2,918
Existing video split	324 train / 171 test; no validation split
Full 75-task graph catalog	`reference_only`; no additional media coverage

Apply and download

Sign in to your own account on the ProAct dataset page, complete the research-access form, and accept all three agreement statements. Applications notify the maintainer in real time at zhuxiaomeng224@gmail.com; each application is reviewed manually. Submission does not automatically grant access. Each researcher, including colleagues in the same group, needs separate approval.

After approval, authenticate with that same account using the official Hugging Face CLI, then download:

hf auth login
hf download Zhuxmmm/ProAct --repo-type dataset --local-dir ProAct

Keep your account and token private. See the public application instructions for the required fields and terms.

Files and frame labels

videos/ contains the masked, silent native MP4s; keyframes/ contains their masked training JPEG derivatives. metadata/ provides portable annotations, video/keyframe indexes, existing splits, task graphs, vocabulary, validation notes and checksums. Join records using video_id and take_id; paths are relative to the dataset root.

Original training filenames are retained. Each JPEG is re-extracted from the final masked native video at the actual decoded ordinal recorded in metadata/keyframes.jsonl, then resized to 768×432. Use frame_index_zero_based for the actual ordinal and filename_frame_label for the historical filename label; they can differ.

Of the 18,747 alignments, 12,727 have strong numeric matches and 6,020 use disclosed historical inference. There are 5,684 preceding-frame selections, including the six historical endpoint labels that select the last available frame N−1. Those six are included in 5,684. Privacy masks and recompression change pixels; filename preservation does not imply pixel identity with the historical JPEGs.

Research limitations

Source task-graph/vocabulary label discrepancies and conflicting embedded camera IDs are preserved and flagged in metadata/validation_notes.json. Original labels, timestamps and split assignments are retained. Historical keyframe alignments marked as inference remain uncertain and should be considered when selecting examples or evaluating results.

External source media

COIN, Ego-Exo4D and UCF-Crime media are obtained separately from their official sources and are not reuploaded in this release. Follow each source's own access requirements and terms. ProAct approval does not grant access to those datasets.

Source	Official access and download instructions
COIN	Dataset website
Ego-Exo4D	Getting Started
UCF-Crime	Official project page

Use and contact

The Data Use Agreement governs the data; the code's Apache 2.0 license is separate. Research results and aggregate metrics may be published under the agreement. Do not re-identify people, reverse privacy transformations, redistribute restricted content or share access credentials. For access questions, written exceptions or private privacy reports, contact zhuxiaomeng224@gmail.com.

Downloads last month: 126

Total file size:

123 GB