EventActivityNet v1.0

EventActivityNet is a generated event voxel tensor dataset derived from ActivityNet videos together with ActivityNet Captions annotations. It provides two alternative temporal groupings over the same canonical 3,263-video set. These are generated tensors, not native event-camera recordings.

Representation	Public path	Videos	Train / validation	Shards (train / validation)	Event shape	Canonical HDF5 bytes
5-bin	data_5bin/	3,263	2,316 / 947	157 / 62	(T5, 5, H, W)	4,355,745,895,245
9-bin	data_9bin/	3,263	2,316 / 947	157 / 62	(T9, 9, H, W)	4,214,122,096,103

The repository contains approximately 8.57 TB of tar-packaged payload. The representations use identical video membership, split assignment, and shard membership. Neither representation is presented as inherently better than the other.

Dataset Structure

Each ActivityNet video corresponds to exactly one HDF5 member in each representation. Train/validation and Large/Medium/Small membership are defined by manifests; the nested scales do not duplicate payload files.

data_5bin/{train,validation}/
data_9bin/{train,validation}/
metadata/{5bin,9bin}/
metadata/video_metadata.jsonl
annotations/
scales/
docs/

Data Format

Every HDF5 file contains exactly:

• events: (T_B, B, H, W), int16;
• voxel_event_start: (T_B,), int64;
• voxel_event_count: (T_B,), int32.

Here B is 5 or 9. For N decoded source frames:

T_B = ceil((N - 1) / B)

Each event slice represents one adjacent decoded-frame transition. An events[t] tensor groups up to B consecutive transition slices. The final group may be partial; unused bins are zero-filled. Timing uses each video's released rational source FPS metadata.

See Dataset Format for schema, timing, and memory-safe loading details.

Release Scales

Scale	Videos	Hours	Train	Validation	Classes	Event-friendly
Large	3,263	106.94	2,316	947	200	65.31%
Medium	1,537	50.00	1,074	463	200	64.80%
Small	667	20.00	473	194	200	64.62%

Small is a strict subset of Medium, and Medium is a strict subset of Large.

Included Metadata

• annotations/activitynet_captions.json: timestamped ActivityNet Captions descriptions;
• annotations/activitynet_actions.json: ActivityNet v1.3 temporal actions;
• annotations/eventactivitynet_alignment.json: derived caption/action alignment;
• annotations/annotation_issues.jsonl: known source annotation quirks;
• metadata/video_metadata.jsonl: shared original-rate timing metadata and per-representation tensor metadata;
• metadata/5bin/ and metadata/9bin/: representation-specific shard manifests, summaries, and checksums;
• scales/: Large, Medium, and Small manifests and public statistics.

Intended Uses

The dataset supports research on generated event voxel representation learning, activity recognition, caption-aligned activity modeling, and comparison across nested dataset scales or temporal groupings.

It is out of scope for identifying people, biometric recognition, surveillance deployment, or consequential decisions about individuals.

Limitations

• The event voxel tensors are generated from conventional videos rather than recorded by an event camera.
• Source FPS and spatial resolution vary by video.
• Timing is based on decoded frame order and released rational nominal or average FPS; per-frame presentation timestamps are not consumed.
• The subset is curated rather than an unbiased conversion of all ActivityNet videos.

Checksums

sha256sum -c metadata/5bin/shard_checksums.sha256
sha256sum -c metadata/9bin/shard_checksums.sha256

Documentation

• Dataset Card
• Dataset Format
• Dataset Generation
• Release Notes
• License Notes

Licensing and Citation

ActivityNet and ActivityNet Captions source terms, licenses, citation obligations, and redistribution restrictions still apply. See License Notes and CITATION.cff.

Release Status

Both complete representations passed final integrity and remote-layout audits. The public payload contains 438 tar shards: 219 per representation.