title: VideoGLaMM ActivityNet Object Review · 20
emoji: 🧩
colorFrom: blue
colorTo: indigo
sdk: static
app_file: index.html
VideoGLaMM ActivityNet Entities GCG — object text review, 20 samples
Independent review Space; the reference VIPOSeg Space is unchanged. Layout adapted from VIPOSeg Grounded Caption Review.
20 distinct original ActivityNet video IDs, one randomly selected eligible segment per video (seed 20261030). The supplied media are 10 sampled RGB frames per segment, not complete videos. The source masks/bounding boxes are supplied on one reference frame only. No mask propagation or new segmentation was performed. Frame time labels are estimates from the source sampling procedure; original video FPS is unknown. Playback is a slideshow at a user-facing inspection rate, not original speed.
Uses the exact words of refined_caption after stripping source phrase/SEG markup and normalizing whitespace. Original refined_caption, new_caption, original JSON, model JSON, derived spans, and mask provenance remain available. All text groups are retained, including groups without masks.
All 20 OpenAI Batch requests completed, with no API/parse failures. GPT saw 10 clean RGB frames plus the all-mask ID overlay, identity sheet, and binary-mask sheet for each sample. 107 entity groups and 188 mentions, 32 source masks. Model candidates: 24 matched, 52 missing mask, 21 not visible in reference, 7 mask quality issue, 2 ambiguous, 1 requires mask union. No complete correspondence candidates; no human-reviewed samples in the immutable source data. Structural validation does not prove semantic correctness or full entity coverage.
The provided annotation guideline is preserved verbatim in data/TEXT_ANNOTATION.original.md and embedded in data/annotation_prompt.txt. Source example inconsistencies and approved execution notes are recorded separately. Original annotation rule SHA256: d6ebb95f5bc1912569660ddf7622329204c71245050d044813e52ac21d425ac6.
Review status and notes are saved only in this browser (versioned localStorage key videoglamm-activitynet-object20-review-v1). They do not change original data or create shared server annotations. Export/import JSON to move reviews between browsers or send them for aggregation. Exports bind to the immutable dataset SHA256. A review mark does not make missing masks complete.
Sources: VideoGLaMM dataset instructions; original ActivityNet Entities GCG archive provided there. Media rights remain with original data owners. This Space is a research annotation preview.