suryadv/concor-video-data-filtered

ConCor Video Data (Filtered)

ConCor Video converts referring-video annotations into correspondences between exact text spans and instance tracklets. This repository publishes the verifier-ready outputs generated for Ref-YouTube-VOS and ReVOS.

The repository name reflects the intended final state. It currently contains the generated training data and the unfiltered evaluation data. eval/filtered/ is reserved for the manual-review decisions and resulting filtered evaluation Parquet that will be uploaded after review.

Repository layout

Path	Contents	Rows	Status
train/verification.parquet	Combined Ref-YouTube-VOS and ReVOS training data	42,165	Generated; not manually filtered
train/manifest.json	Training export counts and provenance	—	Final generation manifest
eval/unfiltered/verification.parquet	Combined validation/test data available from Ref-YouTube-VOS and ReVOS	7,918	Awaiting/completing manual review
eval/unfiltered/manifest.json	Evaluation export counts and provenance	—	Final generation manifest
eval/filtered/decisions.json	Checkpointed manual-review decisions	—	To be uploaded
eval/filtered/verification.parquet	Evaluation rows after applying the decisions	—	To be uploaded

Every row retains its dataset and split fields, so Ref-YouTube-VOS/ReVOS and their original splits can be separated losslessly.

Data generation

The data was generated with the ConCor Video Data Processing pipeline. The checkpointable pipeline preserves a public target tracklet when available, extracts independently visible context entities, uses SAM 3.1 Object Multiplex to segment and track additional entities, validates text-span/tracklet links in both directions, and exports one nested verifier-ready row per referring expression.

The two Parquets are the compact campaign exports. They contain the annotations, tracklet masks, exact span links, extraction/audit metadata, and source-media references required by the verification interface. Original video frames are not bundled; obtain the underlying Ref-YouTube-VOS and ReVOS media under their respective terms and supply that directory as --media-root.

Generation completed without recorded errors:

Export	Expressions	Tracklets	Span links
Training	42,165	129,273	81,867
Evaluation	7,918	24,663	14,394

Manual verification and filtering

The same GitHub project includes the local ConCor video verification interface. It presents all instructions for one video together and supports video-level accept/reject decisions, text edits, exact span-to-tracklet linking, tracklet/span deletion, instruction discard/restore, and decision resumption.

Run it from an installed checkout of the processing project:

concor-video verify \
  --parquet eval/unfiltered/verification.parquet \
  --media-root /path/to/concor-video \
  --decisions eval/filtered/decisions.json \
  --output eval/filtered/verification.parquet \
  --port 8000

Then open http://127.0.0.1:8000. The interface checkpoints edits in browser local storage and can also save decisions.json atomically. Save + export both files writes/downloads both the decisions and the updated Parquet. The unfiltered source Parquet is never modified.

To resume or regenerate the final evaluation table, pass the saved eval/filtered/decisions.json back through the same verification command with eval/unfiltered/verification.parquet as the input. Upload the resulting two files directly into eval/filtered/.

Parquet schema

Each verifier-ready row includes dataset/split identity, video and expression identifiers, text, negative/target provenance, frame references, nested tracklets and masks, span groups and bidirectional links, extraction metadata, SAM prompt audit information, and runtime/disposition fields. JSON-valued nested fields are serialized as strings for portable Parquet/browser handling.

Data sources

This is a derived annotation release for Ref-YouTube-VOS and ReVOS. Consult and comply with the original datasets' access, citation, and licensing terms when obtaining or using their media and annotations.