PointRefiner: per-instance evaluation results and supplementary checkpoints

Supplementary artifacts for the paper "PointRefiner: Sub-Million Parameter Point-Prompted Instance Segmentation" by Bulzan and Cernăzanu-Glăvan (under review at Neurocomputing, 2026). ▎ Contents: ▎ ▎ 1. Per-instance evaluation result JSONs (11 files) for the LVIS v1 val benchmark (187,447 instances, single-click ▎ protocol). Each file contains a summary block (mean IoU, success-at-IoU rates, mean box IoU, timing) and per-sample ▎ ious / box_ious arrays, indexed in the order of the shared click list (clicks_lvis_val_min10_seed0.json, distributed ▎ via the project repository). Methods covered: ▎ - PointRefiner-Large (paper headline) and PointRefiner-Base ▎ - PointRefiner-Large with the v2 inference recipe (hflip TTA, ScaleNet top-2 fallback, selective auto-refinement) ▎ - SAM 1 ViT-B, SAM 2.1-Tiny, SAM 2.1-Large, MobileSAM, EdgeSAM, EfficientSAM-Ti ▎ 2. Smaller-variant ScaleNet checkpoints used in the development-history experiments described in ▎ experiments/README.md of the project repository (not the canonical paper checkpoints — those ship in the repo ▎ itself). ▎ ▎ The trained model weights for the canonical paper system (PointRefiner-Large + ScaleNet-Base, 5.2 MB total) are ▎ released directly in the source repository: https://github.com/AndreiBulzan/pointrefiner. This Zenodo record ▎ complements the repository with the larger per-instance result arrays. ▎ ▎ All evaluations use the LVIS v1 val annotations on top of COCO 2017 images. The shared click list and full ▎ reproduction commands are in the source repository.