dingxingdi Skill: long-video-summarization
1. Capability Definition & Real Case
- Professional Definition: The capability to compress long-form or heavily fragmented multi-scene video into a faithful global account and perform high-level narrative reasoning by integrating evidence across distinct temporal segments. This preserves major events, cross-event logical plot progression, and outcomes without collapsing into vague topical language or omitting decisive developments and situational shifts, even amidst rapid pacing.
- Dimension Hierarchy: Long-Horizon Reasoning->Global Content Integration->long-video-summarization
Real Case
[Case 1]
- Initial Environment: A 100-minute football match video with multiple attacks, substitutions, and momentum swings. The user asks about the second half rather than about one isolated play.
- Real Question: What happens in the second half of the game?
- Real Trajectory: Segment the second half into major phases, identify turning-point actions, suppress low-impact repetitions, and compose a chronologically ordered summary that preserves the decisive score-changing events.
- Real Answer: As the game progressed towards its final stages, Liverpool's offensive pressure gradually intensified. A free kick from Liverpool leading to an own goal in the penalty area later changed the game's momentum, Liverpool once again scored another goal, resulting in a 2-0 victory.
- Why this demonstrates the capability: The answer requires broad coverage and selection of high-salience developments across a long interval. This tests whether the agent can preserve temporal progression and outcome while compressing large amounts of video evidence.
[Case 2]
- Initial Environment: A 90-minute cinematic video featuring multiple segments of astronauts preparing for and executing a technical mission. The first 10 minutes establish the mission's scope on Earth, while the final 20 minutes show the execution phase.
- Real Question: Why are the men dressed in white space suits walking down a ramp at the beginning of the event from 76.0 to 135.0 seconds?
- Real Trajectory: Scan the initial frames to identify the mission context (e.g., preparing for a space mission). Locate the specific temporal window (76s-135s) within the hour-long stream. Correlate the visual of the men on the ramp with the mission goals identified at the very start of the video. Deduce that the action is a procedural prerequisite (pre-launch walk) for the goal established in the long-term context.
- Real Answer: They are preparing for a space mission.
- Why this demonstrates the capability: This demonstrates 'Within-Event to Global Alignment.' The agent matches a specific, time-localized action at the beginning of a sequence to the overarching narrative goal established in the long-form video profile, proving it can ground reasoning in structural plot intent.
[Case 3]
- Initial Environment: A 120-second cinematic advertisement featuring a woman in a pink dress running through a busy cityscape, transitioning rapidly through multiple high-budget locations like New York's Times Square, an apartment, and a taxi cab.
- Real Question: Which classic film shares the same storyline as this fast-paced video? (A) Roman Holiday, (B) The City of Love, (C) The Devil Wears Prada, (D) Before Sunrise.
- Real Trajectory: Perform narrative synthesis across rapid scene cuts. First, identify the 'celebrity runaway' archetype in the first 20s. Next, note the cab interaction with a man who doesn't recognize her at 45s. Finally, detect her return to her career at 110s. By correlating this high-density condensed plot to film history, the agent identifies the core escapist romance theme.
- Real Answer: Roman Holiday.
- Why this demonstrates the capability: This demonstrates 'Cross-Scene Narrative Synthesis' across high-density edits. Bridging rapidly disjointed sequences to extract a unified thematic arc confirms the capability to integrate massive plot gaps into a coherent global scenario.
Pipeline Execution Instructions
To synthesize data for this capability, you must strictly follow a 3-phase pipeline. Do not hallucinate steps. Read the corresponding reference file for each phase sequentially:
Phase 1: Environment Exploration Read the exploration guidelines to discover raw knowledge seeds:
references/EXPLORATION.mdPhase 2: Trajectory Selection Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
references/SELECTION.mdPhase 3: Data Synthesis Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
references/SYNTHESIS.md