HINT-3D: Human-in-the-Loop Interactive Test-Time Adaptation for 3D Segmentation

We present HINT-3D, a human-in-the-loop test-timeadaptation framework for 3D semantic segmentation. A fewcorrective clicks are converted into region masks by a promptable3D interface (PointSAM). These masks supervise stability-awareupdates to a pretrained backbone at inference. We persist theupdates so later scenes start from improved weights, enablingcumulative learning. The wrapper is backbone-agnostic: it requiresonly logits, a mask-to-index bridge, plus access to a small trainableparameter set; we instantiate it on KPConv, RandLA-Net, andPoint Transformer v1. On S3DIS Area-5, HINT-3D deliversstrong effort-accuracy gains within a scene, consistent zero-clickimprovements across scenes, and reduced Expected CalibrationError (ECE), while maintaining responsiveness with head-onlyupdates and uncertainty-gated training. We report mIoU versussaved masks, cross-scene transfer, ECE, latency, and class-specificcorrections on common indoor failure modes.