dghadiya/T2AV_TemporalConstraint_HDEPIC

HD-EPIC real-video benchmark for T2AV judge evaluation

Clips are cut from HD-EPIC (Perrett et al., "HD-EPIC: A Highly-Detailed Egocentric Video Dataset", CVPR 2025). For research use, follow the original HD-EPIC license and cite the paper.

What this is

Real kitchen clips used as "correct" videos. Prompts are written from the HD-EPIC narration and sound annotations, then one axis is broken on purpose to make FAIL items. Each item = one clip + one prompt + a GT label per axis (A, V, AVB, TC-Rel, TC-Cnt). CC is not scored on real videos.

194 original windows, 1,523 items in total:

Variant	Folder suffix	Broken axis	n	How it is made
original	_pass	none (all PASS)	194	original clip and prompt
mute	_mute	A	194	audio removed
absent-sound claim	_A	A	194	prompt claims a sound that is not there
hand swap	_Vhand	V	134	prompt and event list use the other hand
noun swap	_Vnoun	V	194	prompt and event list use another object
audio shift early	_AVBearly	AVB	124	audio track shifted earlier
audio shift late	_AVBlate	AVB	101	audio track shifted later
order swap	_TCR	TC-Rel	194	order of the two events reversed in the prompt
count off	_TCC	TC-Cnt	194	required count changed by +1

Variants that only change the prompt (A, Vhand, Vnoun, TCR, TCC) reuse the original clip, so the same video appears in several folders.

GT hierarchy

GT follows the same hierarchy the judge uses, at event level:

• A FAIL on an event → AVB FAIL for that event
• V FAIL → AVB FAIL only if the broken event is the sounding anchor (36 of 134 hand swaps, 0 noun swaps)
• A, V or AVB FAIL → TC-Rel and TC-Cnt FAIL

broken_event_idx / broken_event_kind columns in the GT csv tell which event was changed.

Files

• videos/<item_id>/<item_id>_v1_raw.mp4 : one clip per item (1,523), same layout as the judge input
• hdepic_medium_prompts.json : prompts in the new_eval.py input format
• hdepic_medium_gt.csv : per-axis GT (hierarchical, use this one)
• items_hdepic_medium_with_clips_v2.json : source items with events, narration timestamps and variant info

For AVB tools

Audio shift items (225) are the real AVB negatives. Originals (194) are the positives for false alarm rate. The Gemini judge (single agent, no tools) detects early shifts 13% and late shifts 32% of the time, so this is where an AVB tool can show value.