Skip to content

Open-source AI podcast clipper. Generate vertical clips with face tracking and burned-in captions. CLI, MCP server, web app.

Pool highlight detection across a folder of clips

Nika Siradzec2c48ed8279c
diff --git a/backend/main.py b/backend/main.py
index 09fe114..98a45a9 100644
--- a/backend/main.py
+++ b/backend/main.py
@@ -329,22 +329,33 @@ def handle_analyze_energy(task_id: str, params: dict):
def handle_detect_highlights(task_id: str, params: dict):
- """Detect highlight clips from a video's fused signal curve (party/action profiles)."""
- from services.saliency import detect_highlights
+ """Detect highlight clips from a video's fused signal curve (party/action profiles).
+ Accepts a single `video_path`, or a list of `video_paths` to pool and rank
+ highlights across a whole folder of clips.
+ """
+ from services.saliency import detect_highlights, detect_highlights_pooled
+
+ video_paths = params.get("video_paths")
video_path = params.get("video_path", "")
- if not video_path:
- emit_result(task_id, "error", error="video_path is required")
+ if not video_paths and not video_path:
+ emit_result(task_id, "error", error="video_path or video_paths is required")
return
- clips = detect_highlights(
- video_path=video_path,
+ common = dict(
profile_name=params.get("profile", "party"),
- top_n=int(params.get("top_n", 8)),
min_dur=float(params.get("min_dur", 8.0)),
max_dur=float(params.get("max_dur", 60.0)),
progress_callback=lambda pct, msg: emit_progress(task_id, "detecting", pct, msg),
)
+ if video_paths:
+ clips = detect_highlights_pooled(
+ video_paths=video_paths, top_n=int(params.get("top_n", 15)), **common
+ )
+ else:
+ clips = detect_highlights(
+ video_path=video_path, top_n=int(params.get("top_n", 8)), **common
+ )
emit_result(task_id, "success", data={"clips": clips, "count": len(clips)})
diff --git a/backend/services/saliency.py b/backend/services/saliency.py
index ed300c2..bcedb8c 100644
--- a/backend/services/saliency.py
+++ b/backend/services/saliency.py
@@ -243,3 +243,39 @@ def detect_highlights(
if progress_callback:
progress_callback(100, f"Found {len(clips)} highlights")
return clips
+
+
+def detect_highlights_pooled(
+ video_paths: list[str],
+ profile_name: str = "party",
+ top_n: int = 15,
+ min_dur: float = 8.0,
+ max_dur: float = 60.0,
+ progress_callback: Optional[Callable] = None,
+) -> list[dict]:
+ """
+ Detect highlights across many videos and rank them globally — "the best N bits
+ from tonight" across a folder of party clips.
+
+ Each returned clip carries a `source_file`. Ranking is reaction-first, then by
+ score, so a genuine laugh in any file outranks a merely loud moment in another.
+ """
+ pooled: list[dict] = []
+ n = len(video_paths) or 1
+ for idx, path in enumerate(video_paths):
+ clips = detect_highlights(
+ path, profile_name=profile_name, top_n=top_n, min_dur=min_dur, max_dur=max_dur
+ )
+ for c in clips:
+ c["source_file"] = path
+ pooled.extend(clips)
+ if progress_callback:
+ progress_callback(
+ int((idx + 1) / n * 100), f"{path}: {len(clips)} highlights"
+ )
+
+ pooled.sort(
+ key=lambda c: (1 if "reaction" in c.get("reasons", []) else 0, c.get("score", 0)),
+ reverse=True,
+ )
+ return pooled[:top_n]
diff --git a/tests/test_saliency.py b/tests/test_saliency.py
index 91f857e..8734ec7 100644
--- a/tests/test_saliency.py
+++ b/tests/test_saliency.py
@@ -110,5 +110,38 @@ class WindowForPeakTests(unittest.TestCase):
self.assertGreaterEqual(start, 0.0)
+class PooledTests(unittest.TestCase):
+ def test_pools_across_files_reaction_first_with_source(self):
+ fake = {
+ "a.mp4": [
+ {"start_second": 10, "end_second": 18, "score": 30.0, "reasons": ["energy_peak"]},
+ ],
+ "b.mp4": [
+ {"start_second": 5, "end_second": 15, "score": 12.0, "reasons": ["reaction"]},
+ ],
+ }
+ orig = sal.detect_highlights
+ sal.detect_highlights = lambda path, **kw: [dict(c) for c in fake[path]]
+ try:
+ pooled = sal.detect_highlights_pooled(["a.mp4", "b.mp4"], top_n=5)
+ finally:
+ sal.detect_highlights = orig
+ self.assertEqual(len(pooled), 2)
+ # reaction outranks the higher-scored energy peak across files
+ self.assertEqual(pooled[0]["reasons"], ["reaction"])
+ self.assertEqual(pooled[0]["source_file"], "b.mp4")
+ self.assertTrue(all("source_file" in c for c in pooled))
+
+ def test_top_n_caps_pool(self):
+ fake = [{"start_second": i, "end_second": i + 8, "score": float(i), "reasons": ["energy_peak"]} for i in range(10)]
+ orig = sal.detect_highlights
+ sal.detect_highlights = lambda path, **kw: [dict(c) for c in fake]
+ try:
+ pooled = sal.detect_highlights_pooled(["a.mp4"], top_n=3)
+ finally:
+ sal.detect_highlights = orig
+ self.assertEqual(len(pooled), 3)
+
+
if __name__ == "__main__":
unittest.main()