--- pretty_name: V2-SAM EgoExo FullTrain image closure language: - en task_categories: - image-segmentation tags: - ego-exo4d - ego2exo - exo2ego - research-artifact license: other --- # V2-SAM Ego2Exo/Exo2Ego FullTrain image closure Research repack for the V2-SAM Ego2Exo and Exo2Ego training pipelines. It contains the two FullTrain annotations and exactly the JPEG union referenced by them. This card documents packaging integrity; it does not grant additional rights to the source data. Users remain responsible for complying with the underlying Ego-Exo4D terms and any redistribution requirements. ## Why this repack exists The earlier local `Ego-Exo4D-Relation-Train-Mini` extraction was not a proven closure of both training directions. Auditing the union of the two FullTrain JSON files found 6,431 referenced paths missing before repair. Training with that snapshot could fail late, silently omit samples in ad-hoc loaders, or make Ego2Exo and Exo2Ego experiments use different image coverage. This repack fixes the data artifact rather than changing annotations: - both official FullTrain JSONs are included unchanged; - referenced image paths are de-duplicated across both directions; - every referenced JPEG is present in the final image root; - every final JPEG passed the image-readability check; - files are split into independent standard tar archives; - file counts, byte sizes, SHA256 hashes, and repair provenance are retained. ## Verified closure | Item | Verified value | | --- | ---: | | Ego2Exo records | 110,118 | | Ego2Exo unique referenced images | 220,236 | | Exo2Ego records | 123,381 | | Exo2Ego unique referenced images | 246,762 | | Union referenced JPEGs | 249,642 | | Missing before repair | 6,431 | | Missing after repair | 0 | | Unreadable JPEGs after repair | 0 | | Episode groups | 754 | | Tar shards | 4 | The two annotation files retain these identities: | File | SHA256 | | --- | --- | | `Ego2Exo_FullTrain.json` | `bb83cac92179b21833d2465ce29d5b0cf81668d8443d569279313c891472175a` | | `Exo2Ego_FullTrain.json` | `212e7291990453b087f91dd848dd8a8a8f11a7a0a110fb257be760ca9c8c3058` | `provenance/final_closure.json` is the authoritative closure report. `provenance/repair_receipt.json` records the repair audit. Its `already_present_count=6431` means the final repair rerun found all previously missing paths already restored; it is not a claim that no repair had been needed. ## Repository layout ```text . ├── MANIFEST.json ├── SHA256SUMS ├── README.md ├── data/ │ ├── data_segswap-00000.tar │ ├── data_segswap-00001.tar │ ├── data_segswap-00002.tar │ └── data_segswap-00003.tar ├── json/Ego-Exo4D-Seg/frame-level/ │ ├── Ego2Exo_FullTrain.json │ └── Exo2Ego_FullTrain.json └── provenance/ ├── final_closure.json └── repair_receipt.json ``` Each `data/*.tar` is independently readable. Extract all four into the same destination so their path trees merge. ## Download, verify, and extract Authenticate with an account that can read this private dataset: ```bash hf download Travor278/V2SAM-EgoExo-Train-Mini-Complete \ --repo-type dataset \ --local-dir V2SAM-EgoExo-Train-Mini-Complete cd V2SAM-EgoExo-Train-Mini-Complete sha256sum -c SHA256SUMS mkdir -p extracted for shard in data/data_segswap-*.tar; do tar -xf "$shard" -C extracted done ``` Point the V2-SAM dataset `sam2_folder` at the extracted image tree and select the matching JSON for the desired direction: - Ego2Exo: `json/Ego-Exo4D-Seg/frame-level/Ego2Exo_FullTrain.json` - Exo2Ego: `json/Ego-Exo4D-Seg/frame-level/Exo2Ego_FullTrain.json` Do not concatenate the JSONs unless the training code explicitly implements a mixed-direction dataset contract. ## Integrity files - `MANIFEST.json`: shard counts, sizes, hashes, JSON identities, and closure report identities. - `SHA256SUMS`: command-line integrity verification for every payload and metadata file in this repository. - `provenance/final_closure.json`: before/after missing counts and JPEG audit. - `provenance/repair_receipt.json`: source archive and repair-run evidence. The repository is complete only when all SHA256 checks pass and the closure report has `ready=true`, `post_repair_missing_count=0`, and `post_repair_bad_jpeg_count=0`.