= tau] on a held-out slice and filters strictly when the teacher's confident set is as reliable as the confidence it demands (pi_kept >= tau), falling back to a self-adaptive floor when it is not. The boundary is the pre-existing operating threshold, not a value tuned to mIoU — and it makes the correct strict-vs-floor call blind on all six DINOv2 teachers we tested.\n\nIt recovers the UniMatch V2 operating point where the teacher is reliable (Pascal VOC 1/8: 87.4 against 87.9 reported; Cityscapes within 0.5) and improves on it where the confident set is not (ADE20K, pi_kept about 89% rather than 98%: +1.5 mIoU, single seed).\n\nCode, weights for Pascal / Cityscapes / ADE20K, and a Gradio demo you can run locally are in the repo.","html":"<p>Author here.</p>\n<p>Pseudo-label selection rules — dynamic thresholds, per-class curricula, soft confidence weights — were designed for noisy, under-confident ResNet teachers. A DINOv2 teacher changes the regime: confidence saturates (98% of Pascal pixels sit at or above 0.95), so an adaptive cutoff floods the retention mask and self-training decays into confirmation bias. The filtering that helped a weak teacher can hurt a strong one.</p>\n<p>CW-BASS v2 measures the regime instead of assuming it. A one-pass reliability gate estimates pi_kept = Pr[correct | c >= tau] on a held-out slice and filters strictly when the teacher's confident set is as reliable as the confidence it demands (pi_kept >= tau), falling back to a self-adaptive floor when it is not. The boundary is the pre-existing operating threshold, not a value tuned to mIoU — and it makes the correct strict-vs-floor call blind on all six DINOv2 teachers we tested.</p>\n<p>It recovers the UniMatch V2 operating point where the teacher is reliable (Pascal VOC 1/8: 87.4 against 87.9 reported; Cityscapes within 0.5) and improves on it where the confident set is not (ADE20K, pi_kept about 89% rather than 98%: +1.5 mIoU, single seed).</p>\n<p>Code, weights for Pascal / Cityscapes / ADE20K, and a Gradio demo you can run locally are in the repo.</p>\n","updatedAt":"2026-08-14T04:26:38.094Z","author":{"_id":"67c4ec6a7d39ef70f0468f26","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67c4ec6a7d39ef70f0468f26/vBpYHf3zzqbmzn_7aaDl2.jpeg","fullname":"Ebenezer Tarubinga","name":"psychofict","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9123484492301941},"editors":["psychofict"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/67c4ec6a7d39ef70f0468f26/vBpYHf3zzqbmzn_7aaDl2.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.12773","authors":[{"_id":"6a7e930842823931a1f1768e","user":{"_id":"67c4ec6a7d39ef70f0468f26","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67c4ec6a7d39ef70f0468f26/vBpYHf3zzqbmzn_7aaDl2.jpeg","isPro":false,"fullname":"Ebenezer Tarubinga","user":"psychofict","type":"user","name":"psychofict"},"name":"Ebenezer Tarubinga","status":"claimed_verified","statusLastChangedAt":"2026-08-14T08:45:04.700Z","hidden":false}],"publishedAt":"2026-08-13T00:00:00.000Z","submittedOnDailyAt":"2026-08-14T00:00:00.000Z","title":"CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers","submittedOnDailyBy":{"_id":"67c4ec6a7d39ef70f0468f26","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67c4ec6a7d39ef70f0468f26/vBpYHf3zzqbmzn_7aaDl2.jpeg","isPro":false,"fullname":"Ebenezer Tarubinga","user":"psychofict","type":"user","name":"psychofict"},"summary":"Semi-supervised semantic segmentation has long turned on one question, which pseudo-labels to trust, and a generation of selection rules, dynamic thresholds, per-class curricula, soft confidence weights, answered it for the noisy, under-confident ResNet teachers of their day. Self-supervised foundation encoders change the regime: with a DINOv2 teacher, confidence saturates, so the filtering that helped a weak teacher can hurt a strong one. We propose CW-BASS v2, a saturation-aware pseudo-label selection method that reads the teacher's confidence regime rather than committing to one rule. It pairs held-out calibration, an unbiased per-class noise estimate, with a self-adaptive confidence floor that provably bounds retention away from 1, and combines them in a one-pass gate: measure the reliability of the teacher's confident set, pi_kept = Pr[correct | c >= tau], on a held-out slice, and filter strictly when it meets the confidence demanded (pi_kept >= tau), falling back to the adaptive floor otherwise. The boundary is the pre-existing operating threshold, not a value tuned to mIoU, and across six DINOv2 teachers it makes the correct strict-vs-floor call blind. CW-BASS v2 thus recovers the UniMatch V2 operating point on the saturated benchmarks by selecting strict (Pascal VOC 1/8 87.4 against its reported 87.9; Cityscapes within 0.5), and improves on it where the confident set is unreliable (pi_kept ~ 89%, ADE20K), where the floor edges ahead (+1.5 mIoU, single seed). The gate is principled because the failure it avoids is measured, not assumed: on a reliable, saturated teacher the confidence distribution's dynamic range collapses (98% of Pascal pixels >= 0.95), so an adaptive cutoff floods the retention mask and self-training decays into confirmation bias.","upvotes":0,"discussionId":"6a7e930842823931a1f1768f","projectPage":"https://psychofict.github.io/CW-BASS-v2/","githubRepo":"https://github.com/psychofict/CW-BASS-v2","githubRepoAddedBy":"user","ai_summary":"CW-BASS v2 selects pseudo-labels by measuring teacher reliability on held-out data and applying either strict filtering or an adaptive floor to avoid confirmation bias under saturated confidence.","ai_keywords":["semi-supervised semantic segmentation","pseudo-label selection","DINOv2","self-supervised foundation encoders","confidence saturation","held-out calibration","adaptive confidence floor","confirmation bias","mIoU"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":0,"organization":{"_id":"6a4c80d65ff747e27049ee7b","name":"ebenworks","fullname":"Ebenworks Systems","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/67c4ec6a7d39ef70f0468f26/wR4jDgD-kkrWUGxtRqH27.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[],"acceptLanguages":["en"],"organization":{"_id":"6a4c80d65ff747e27049ee7b","name":"ebenworks","fullname":"Ebenworks Systems","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/67c4ec6a7d39ef70f0468f26/wR4jDgD-kkrWUGxtRqH27.jpeg"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.12773.md","query":{}}">
CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers
Abstract
CW-BASS v2 selects pseudo-labels by measuring teacher reliability on held-out data and applying either strict filtering or an adaptive floor to avoid confirmation bias under saturated confidence.
Semi-supervised semantic segmentation has long turned on one question, which pseudo-labels to trust, and a generation of selection rules, dynamic thresholds, per-class curricula, soft confidence weights, answered it for the noisy, under-confident ResNet teachers of their day. Self-supervised foundation encoders change the regime: with a DINOv2 teacher, confidence saturates, so the filtering that helped a weak teacher can hurt a strong one. We propose CW-BASS v2, a saturation-aware pseudo-label selection method that reads the teacher's confidence regime rather than committing to one rule. It pairs held-out calibration, an unbiased per-class noise estimate, with a self-adaptive confidence floor that provably bounds retention away from 1, and combines them in a one-pass gate: measure the reliability of the teacher's confident set, pi_kept = Pr[correct | c >= tau], on a held-out slice, and filter strictly when it meets the confidence demanded (pi_kept >= tau), falling back to the adaptive floor otherwise. The boundary is the pre-existing operating threshold, not a value tuned to mIoU, and across six DINOv2 teachers it makes the correct strict-vs-floor call blind. CW-BASS v2 thus recovers the UniMatch V2 operating point on the saturated benchmarks by selecting strict (Pascal VOC 1/8 87.4 against its reported 87.9; Cityscapes within 0.5), and improves on it where the confident set is unreliable (pi_kept ~ 89%, ADE20K), where the floor edges ahead (+1.5 mIoU, single seed). The gate is principled because the failure it avoids is measured, not assumed: on a reliable, saturated teacher the confidence distribution's dynamic range collapses (98% of Pascal pixels >= 0.95), so an adaptive cutoff floods the retention mask and self-training decays into confirmation bias.
Community
Author here.
Pseudo-label selection rules — dynamic thresholds, per-class curricula, soft confidence weights — were designed for noisy, under-confident ResNet teachers. A DINOv2 teacher changes the regime: confidence saturates (98% of Pascal pixels sit at or above 0.95), so an adaptive cutoff floods the retention mask and self-training decays into confirmation bias. The filtering that helped a weak teacher can hurt a strong one.
CW-BASS v2 measures the regime instead of assuming it. A one-pass reliability gate estimates pi_kept = Pr[correct | c >= tau] on a held-out slice and filters strictly when the teacher's confident set is as reliable as the confidence it demands (pi_kept >= tau), falling back to a self-adaptive floor when it is not. The boundary is the pre-existing operating threshold, not a value tuned to mIoU — and it makes the correct strict-vs-floor call blind on all six DINOv2 teachers we tested.
It recovers the UniMatch V2 operating point where the teacher is reliable (Pascal VOC 1/8: 87.4 against 87.9 reported; Cityscapes within 0.5) and improves on it where the confident set is not (ADE20K, pi_kept about 89% rather than 98%: +1.5 mIoU, single seed).
Code, weights for Pascal / Cityscapes / ADE20K, and a Gradio demo you can run locally are in the repo.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.12773 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.12773 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.