| name | kuavi-predictive |
| description | Predictive video understanding — anticipate actions, predict future content, verify coherence, classify activities |
Predictive Video Analysis
Use V-JEPA 2-powered predictive tools for forward-looking video understanding.
Tools
kuavi_anticipate_action
Predict what happens next after a given timestamp.
- Uses V-JEPA 2 predictor when available, falls back to embedding similarity
- Returns predicted action, confidence, and supporting evidence
kuavi_anticipate_action(time_point=45.0)
kuavi_predict_future
Predict future video content from a time range.
- Temporal continuation using V-JEPA 2 predictor embeddings
- Returns predicted content description with confidence
kuavi_predict_future(start_time=30.0, end_time=45.0)
kuavi_verify_coherence
Score temporal coherence across video segments and detect anomalies.
- Compares predicted vs actual embeddings at segment boundaries
- Flags surprising transitions where prediction diverges from reality
kuavi_verify_coherence()
kuavi_classify_segment
Classify a video segment using attentive probes trained on benchmark tasks.
- Available tasks: SSv2, K400, Diving48, and more
- Returns top-K class labels with confidence scores
kuavi_classify_segment(start_time=10.0, end_time=20.0)
Example Workflows
"What happens next?"
kuavi_search_video("current activity", field="action") to locate the moment
kuavi_anticipate_action(time_point=<end_of_activity>) to predict next action
kuavi_extract_frames around the predicted time to verify
"Is this video coherent?"
kuavi_verify_coherence() to get per-segment coherence scores
- Look for segments with low coherence (anomalies)
kuavi_extract_frames around anomalous transitions to inspect
"Classify this activity"
kuavi_classify_segment(start_time, end_time) for benchmark labels
- Cross-reference with
kuavi_search_video(field="action") for caption-based description