coverage string | pairs int64 | objects int64 | subsets dict | reference_record_hash_mask_and_choice_audit string | historical_witnesses string | frame_and_group_bootstrap_recompute string | compact_sha256 string | frozen_selection_sha256 string | no_target_fitting bool | missing_linked_direction string |
|---|---|---|---|---|---|---|---|---|---|---|
exact | 6,513 | 6,513 | {
"behaviorsim/exo2ego": 243,
"behaviorsim/exo2exo": 222,
"droid/exo2ego": 2162,
"droid/exo2exo": 2792,
"egoexo4d/exo2exo": 1094
} | passed | passed | passed | c9ba43566bf2e227e05749281a75cedfdeb6240d6af6fa999b1a2dcd48fa024b | a913b51b9aa068706870adc070f8ec32e437776cb070c05150853ae896f043c3 | true | Jenk9 linked repo contains no Exo2Ego; separately verified EgoExo4D companion reused. |
V2-SAM PCCS cross-view segmentation results
This public result pack contains the frozen zero-shot evaluation artifacts for the context-consistency PCCS extension of V2-SAM. It covers BehaviorSim, DROID, and selected EgoExo4D Exo2Exo pairs. Raw source images, source-dataset annotations, model checkpoints, and candidate-mask banks are not redistributed here.
The implementation keeps the original PCCS candidate masks fixed and adds contextual evidence around the native object ROI. The two frozen variants use 1.5x and 2.0x object-centered ROIs. All methods in each row below share the same saved candidate masks, pairing, and evaluation pipeline; no target-dataset fitting was performed.
Full-set results
Frame-level IoU is reported in percent. Parentheses show the paired improvement over the unchanged PCCS baseline. Confidence intervals and all additional method variants are in multidata_20260920/full_all_results.json.
| Dataset | Direction | Pairs | Baseline | PCCS + context 1.5x | PCCS + context 2.0x | O-MaMa consensus |
|---|---|---|---|---|---|---|
| BehaviorSim | Exo2Ego | 243 | 56.6129 | 60.9908 (+4.3779) | 60.9894 (+4.3765) | 60.6294 |
| BehaviorSim | Exo2Exo | 222 | 56.9832 | 60.9379 (+3.9547) | 60.3487 (+3.3655) | 61.5733 |
| DROID | Exo2Ego | 2,162 | 68.7233 | 73.3014 (+4.5782) | 73.2746 (+4.5513) | 69.8146 |
| DROID | Exo2Exo | 2,792 | 79.7886 | 82.3027 (+2.5141) | 82.4009 (+2.6123) | 81.9902 |
| EgoExo4D selected | Exo2Exo | 1,094 | 38.5921 | 42.0949 (+3.5028) | 42.7655 (+4.1734) | 42.2977 |
The companion EgoExo4D Exo2Ego result was previously verified on 46,515 pairs / 109,253 objects: 53.9195 baseline, 56.3794 with the 1.5x variant (+2.4599), and 56.4334 with the 2.0x variant (+2.5140). It was not rerun in the multidataset job and is labeled accordingly in the result JSON.
Validation contract
The published full run contains 6,513 evaluated pair/object records across the five newly evaluated dataset-direction splits. validation.json records successful checks for:
- exact coverage and unique keys;
- frozen candidate-mask hashes and choices;
- historical pipeline witnesses;
- frame/group bootstrap recomputation;
- readable compressed compact records.
The evaluation used bootstrap seed 20260920. Every PCCS context variant has a positive paired 95% confidence interval versus baseline on all five newly evaluated splits.
Repository contents
multidata_20260920/: evaluation code, frozen policy definitions, result JSON, compact scalar records, manifests, execution receipts, and experiment logs.reports/FINAL_REPORT.md: consolidated experimental report and interpretation.reports/METHOD.md: method description and visual explanation.RELEASE_MANIFEST.json: byte size and SHA-256 for every published file.
The corresponding source branch is Travor278/EgoExoSeg — pccs-omama-gain-20260916. The evaluated method builds on V2-SAM-O. Source datasets remain governed by their own repositories and terms:
- Travor278/behaviorsim-crossview-seg
- Travor278/droid-crossview-seg
- Jenk9/exo2exo-selected-EgoExo4D-pair-same-random
This repository publishes derived evaluation evidence and experiment code only. Obtain source data and model checkpoints from their original repositories.
- Downloads last month
- 139