JaneDing2025/simact_video

SimAct Video

SimAct Video contains before/after image pairs for short human action segments, converted from multiple video action-recognition datasets into a shared JSONL + tar layout.

Each JSONL row points to two images: precon is the starting visual state and postcon is the ending visual state. Image paths are relative to the repository root and match paths stored inside the tar chunks.

Dataset Summary

This release contains 543,005 examples and 1,086,010 images across 6 datasets. Images are packaged into 35 tar chunks.

Dataset	Train	Validation	Test	Total
ego4d_fho_lta	63,933	33,095	0	97,028
fineaction	57,698	24,218	0	81,916
holoassist	124,752	17,902	0	142,654
hd_epic	0	59,415	0	59,415
assembly101	47,539	15,675	21,893	85,107
epic_kitchens_100	67,217	9,668	0	76,885
Total	361,139	159,973	21,893	543,005

Repository Layout

Files are organized under data/<dataset>/:

data/
  <dataset>/
    train.jsonl
    validation.jsonl
    test.jsonl                  # only when available
    train_chunk_000.tar
    train_chunk_001.tar
    validation_chunk_000.tar
    test_chunk_000.tar          # only when available
    train.done
    validation.done
    test.done                   # only when available

The tar files contain images at paths like:

data/<dataset>/<split>/images/<id>_pre.jpg
data/<dataset>/<split>/images/<id>_post.jpg

The same relative paths are used in the JSONL precon and postcon fields.

JSONL Schema

Each split file is a JSON Lines file. Each line is one example with this schema:

{
  "dataset": "string",
  "split": "train | validation | test",
  "id": "string",
  "original_id": "string",
  "precon": "string",
  "postcon": "string",
  "source": "string",
  "raw_action": {
    "sentence": "string",
    "verb": ["string"],
    "noun": ["string"]
  }
}

Field meanings:

Field	Type	Description
dataset	string	Dataset name used in this release, such as fineaction or epic_kitchens_100.
split	string	Split name: train, validation, or test.
id	string	Release-stable example id generated by this conversion, formatted as <dataset>_<split>_<8-digit-index>.
original_id	string	Original dataset/triplet id before remapping. Use this for tracing back to the source annotations.
precon	string	Repository-relative path to the start-state image.
postcon	string	Repository-relative path to the end-state image.
source	string	Short description of the source annotation file or field used to build the action segment.
raw_action.sentence	string	Natural-language action text from the source dataset when available; otherwise "".
raw_action.verb	list[string]	Source verb label(s) when available; otherwise []. Single-label datasets still use a list.
raw_action.noun	list[string]	Source noun/object label(s) when available; otherwise []. Single-label datasets still use a list.

raw_action preserves whatever the source dataset provides. Some datasets provide only a natural-language label, some provide verb/noun labels, and some provide both. Missing values are represented as an empty string for sentence and empty lists for verb or noun.

Example row:

{"dataset":"hd_epic","split":"validation","id":"hd_epic_validation_00000000","original_id":"P01-20240202-110250-1","precon":"data/hd_epic/validation/images/hd_epic_validation_00000000_pre.jpg","postcon":"data/hd_epic/validation/images/hd_epic_validation_00000000_post.jpg","source":"HD_EPIC_Narrations.pkl narration + verbs/nouns lists","raw_action":{"sentence":"Open the upper cupboard by holding the handle of the cupboard with the left hand.","verb":["open","hold"],"noun":["upper cupboard","handle of cupboard"]}}