Jenk9/exo2exo-selected-EgoExo4D-pair-same-random

Exo2Exo Selected SAM2 Prompt Pair Same+Random

This dataset is a compact Exo2Exo zero-shot segmentation evaluation set derived from an Ego-Exo4D-Seg test subset.

It contains human-selected exocentric prompt masks. Each selected prompt is paired with exocentric target-view GT masks in two temporal modes:

• same_frame: prompt frame and target frame are temporally matched.
• random_mismatch: target frame is randomly sampled from frames where the same object exists.

Contents

Path	Description
annotations/v2sam/exo2exo_selected_sam2_prompt_pair_same_random.json	V2-SAM / V2-SAM-O style annotation JSON
annotations/objectrelator/exo2exo_selected_sam2_prompt_pair_same_random_objectrelator_v2.json	ObjectRelator / PSALM / SEEM style annotation JSON
images/	Referenced image files only, preserving original relative paths
metadata/	Stats generated during dataset construction
scripts/	Construction scripts used for this release
manifest.json	Dataset counts and distribution summary

Statistics

• V2-SAM style samples: 1094
• ObjectRelator/PSALM/SEEM style samples: 1093
• Unique referenced images: 1557
• Image size: 271.7 MiB
• Temporal modes: {'same_frame': 547, 'random_mismatch': 547}
• Takes: 20
• Objects: 79

Usage

For V2-SAM / V2-SAM-O evaluation, set the dataset image root to the downloaded repository's images directory and use:

DATA_ROOT=/path/to/exo2exo-selected-sam2-pair-same-random/images
JSON=/path/to/exo2exo-selected-sam2-pair-same-random/annotations/v2sam/exo2exo_selected_sam2_prompt_pair_same_random.json

For ObjectRelator / PSALM / SEEM evaluation, use:

DATA_ROOT=/path/to/exo2exo-selected-sam2-pair-same-random/images
JSON=/path/to/exo2exo-selected-sam2-pair-same-random/annotations/objectrelator/exo2exo_selected_sam2_prompt_pair_same_random_objectrelator_v2.json

Notes

The prompt masks are pseudo masks generated by V2-SAM and manually selected/filtered, with SAM2 mask-only refinement used in the final selected prompt set. Target masks are GT masks from the Ego-Exo4D-Seg annotation source.