← 研踪 PaperTrailarXiv:2607.13361 · v2
Detector Confidence Signals Presence Rather Than Occlusion in Cluttered Manipulation
首次提交 2026-07-15 08:54
版本更新 2026-10-26 22:17
雷达观测 2026-10-29 00:35
摘要
An open-vocabulary detector's confidence that a named object is present barely changes when clutter hides all but an eighth of it, and can even rise as clutter grows. That score is read as a visibility signal: it thresholds detections, grounds language, and gates when an agent acquires another view. We measure it against a geometry oracle that supplies detector-free ground-truth visibility. As true visibility falls from every scene to one in eight, the confidence stays flat and uncorrelated, while the detector reports the target present in nine of ten scenes by firing on same-category distractors: the score says the category is present somewhere, not that the target is visible. A confidence metric understates the benefit of resolving occlusion tenfold ($8$ against $88$ points), and a confidence gate stays put exactly when the object is hidden. A single camera motion makes the target localizable in every occluded scene, against $5\%$ from the fixed view. The same identity failure holds across three detectors, nine object categories, two simulators, and real video, where it calls the object present in $99\%$ of occluded frames, with its box off the target. We release the benchmark and recommend target-grounded signals for gating and evaluation.
作者备注：Accepted to IEEE MSN 2026. 9 pages
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v3 · 2026-10-29 20:45 · 最新观测版本
Detector Confidence Signals Presence Rather Than Occlusion in Cluttered Manipulation
Detector confidence is often used to decide whether a robot needs another view of a target. We test this signal against detector-free visibility from geometry segmentation. In controlled clutter, Grounding DINO confidence remains nearly constant as target visibility falls from all scenes to one in eight. OWLv2 and SAM 3 retain positive correlations with visibility, but all three detectors lose target-instance localization under heavy occlusion. Grounding DINO reports a hidden target present in about nine of ten scenes by detecting same-category distractors. Consequently, confidence understates the benefit of additional views ($8$ versus $88$ percentage points). In a $40$-scene localization comparison, additional-view selection and fusion increase target localizability from $5\%$ to $100\%$. Experiments cover nine categories, two simulators, and real video. On DAVIS, confidence correlates positively with visible area, yet $99\%$ of occluded frames yield a detection and $94\%$ of those detections miss the tracked target. These results distinguish category presence from target visibility and motivate target-grounded evaluation of active perception.
查看此版本解读 →查看此版本原文 ↗v2 · 2026-10-26 22:17
Detector Confidence Signals Presence Rather Than Occlusion in Cluttered Manipulation
An open-vocabulary detector's confidence that a named object is present barely changes when clutter hides all but an eighth of it, and can even rise as clutter grows. That score is read as a visibility signal: it thresholds detections, grounds language, and gates when an agent acquires another view. We measure it against a geometry oracle that supplies detector-free ground-truth visibility. As true visibility falls from every scene to one in eight, the confidence stays flat and uncorrelated, while the detector reports the target present in nine of ten scenes by firing on same-category distractors: the score says the category is present somewhere, not that the target is visible. A confidence metric understates the benefit of resolving occlusion tenfold ($8$ against $88$ points), and a confidence gate stays put exactly when the object is hidden. A single camera motion makes the target localizable in every occluded scene, against $5\%$ from the fixed view. The same identity failure holds across three detectors, nine object categories, two simulators, and real video, where it calls the object present in $99\%$ of occluded frames, with its box off the target. We release the benchmark and recommend target-grounded signals for gating and evaluation.
查看此版本解读 →查看此版本原文 ↗