Upload object detection v2 artifacts

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+detections.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,47 @@
----
-license: mit
----
+# Object detection
+
+This directory contains concepts and object detections generated from the second keyframe set. The primary artifacts are `concepts.jsonl` and `detections.jsonl`; `artifact_metadata.jsonl` records the input, model, prompt version, subset, and Git lineage.
+
+## `concepts.jsonl`
+
+Each line is one `ConceptRecord`, a short list of general concepts proposed for one keyframe.
+
+Common fields:
+
+- `concept_id`: stable concept identifier
+- `video_id`: source video identifier
+- `time_span`: `[timestamp_ms, timestamp_ms]` in source video time
+- `keyframe_id`: referenced canonical keyframe
+- `concepts`: unique general visual concepts used as detection prompts
+
+## `detections.jsonl`
+
+Each line is one `TrackingDetectionRecord`, a detected object in one keyframe.
+
+Common fields:
+
+- `detection_id`: stable observation identifier
+- `video_id`: source video identifier
+- `time_span`: `[timestamp_ms, timestamp_ms]` in source video time
+- `label`: detected object class
+- `bbox`: normalized object region with `x_min`, `y_min`, `x_max`, and `y_max`
+- `confidence`: SAM3 detection score
+- `concept_id`: concept that prompted this detection
+- `track_id`: sparse association within the source shot
+- `frame_index`: source frame index
+
+The published files are merged across all 700 videos. Each JSONL line keeps its `video_id`, so consumers can filter one video without restoring the former shard directories.
+
+## Generation flow
+
+A vision endpoint proposes general concepts from the keyframes. SAM3 verifies and localizes those concepts, then associates matching boxes across adjacent keyframes in the same shot. It publishes concepts and detections separately so visible text remains the responsibility of OCR.
+
+Run detection for the GEBD and MaxInfo keyframe set:
+
+```bash
+make data-detection \
+  JOB='concepts detection' \
+  INPUT=/data/aic/shared/artifacts/keyframesv2/keyframes.jsonl \
+  OUTPUT=/data/aic/shared/artifacts/object_detectionv2 \
+  SUBSET=700vid
+```

artifact_metadata.jsonl

@@ -0,0 +1,2 @@
+{"artifact":"concepts.jsonl","input":"../../keyframesv2/keyframes/00cdb8ca10c069f8","subset":"700vid","stage":"concept_extractor.keyframe","model":"Cosmos3-Super","prompt_version":"aic26_keyframe_entities_v6","git_commit":"598401c15a646a0586135a1a380eadad63a89d2d"}
+{"artifact":"detections.jsonl","input":"concepts.jsonl","subset":"700vid","stage":"object_detection.sam3_image","model":"facebook/sam3","prompt_version":"multi_query_keyframe_v1","git_commit":"598401c15a646a0586135a1a380eadad63a89d2d"}

detections.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bbe29f5e80fc3ed2cdfac6cce466947d73e05b11e83739ab35b5eeb5190b56c
+size 136616392