A Matched Intervention Benchmark and Evaluation Tool for Auditing Internal Global-Feature Ablation in Social Relationship Recognition

Image benchmark and CPU evaluation resources accompanying the WACV 2027 paper A Matched Intervention Benchmark and Evaluation Tool for Auditing Internal Global-Feature Ablation in Social Relationship Recognition, by Ichiro Tateyama, Matus Tanonwong, and Koichi Hashimoto. The archive contains 218 PISC source photographs, 872 matched edits (four per source), 31 additional L3 images for a separate invalid-edit test, and 194 distinct protection masks. All scored image files retain their original encoded bytes; no resizing or recompression is applied for distribution. The 31 L3 images are excluded from the primary 872-image analysis. The evaluation resources include 31 ConSoR/GRIT configurations, 27,032 matched prediction records, 10,246 original/ablation records, source-level metrics, bootstrap analysis, L3 measurements, and a CPU evaluator. The image manifest links image IDs, conditions, target pairs, geometry, preservation contracts and SHA-256 hashes to the predictions. Preservation is not uniform: 773 matched edits have source-pixel-exact protection, and 99 use semantic preservation or local reconstruction. We provide sensitivity analyses on fixed subsets. We also conducted a human review of all 872 images, independently of the VLM reviews; the human reviewer did not see the VLM judgments. Model reviews and disagreements are documented; there was no external two-human validation study. These interventions measure bounded behavioral responses and do not establish causal understanding. The PISC source dataset is by Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S. Kankanhalli (doi:10.5281/zenodo.1059155). Its official record lists CC BY 4.0. Changes include generated replacement settings, protected-foreground compositing where used, and reviewed local repairs. Attribution and modification notices are included. The evaluation software carries the MIT license supplied in evaluation/LICENSE. Model weights and API credentials are not included. Download layout: The image payload is supplied as 32 ordinary ZIP archives plus pisc_metadata_and_evaluation.zip. Download all parts and extract them into the same directory. DOWNLOAD.md gives instructions, and SHA256SUMS.txt verifies the downloadable files. All image, mask, prediction and statistic bytes are preserved. The supplementary PDF has a formatting-only correction for readability. Both manuscript PDFs use PDF 1.4 and Web optimization; the main L3 illustration retains the original image resolution. The parts are a distribution layout.