Research map / Detection and segmentation / Weakly supervised detection and segmentation
Diverse Learner: Exploring Diverse Supervision for Semi-Supervised Object Detection
Linfeng Li, Minyue Jiang, Yue Yu, Wei Zhang, Xiangru Lin, Yingying Li, Xiao Tan, Jingdong Wang, Errui Ding
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-threshold classification loss (MTC) (ablation) | STAC baseline with box jitter · MS-COCO val (1% labeled) | mAP: 20.6 -> 22.1 | Improved |
| add pseudo label-aware erasing (PLAE) on top of MTC (ablation) | Baseline + MTC · MS-COCO val (1% labeled) | mAP: 22.1 -> 22.7 | Improved |
| add diverse learner (two cross-supervised teacher-student pairs) on top of MTC and PLAE (ablation) | Baseline + MTC + PLAE · MS-COCO val (1% labeled) | mAP: 22.7 -> 23.7 | Improved |
| add conventional mutual learning instead of DL (ablation) | Baseline + MTC + PLAE · MS-COCO val (1% labeled) | mAP: 22.7 -> 19.0 | Hurt |
| use KL divergence loss rather than ignoring uncertain pseudo labels (ablation) | Baseline + PLAE · MS-COCO val (1% labeled) | mAP: 21.2 -> 22.7 | Improved |
| use KL divergence loss rather than binary cross entropy for uncertain pseudo labels (ablation) | Baseline + PLAE · MS-COCO val (1% labeled) | mAP: 22.2 -> 22.7 | Improved |
| lower the MTC lower threshold from 0.6 to 0.8 (ablation) | Diverse Learner · MS-COCO val (1% labeled) | mAP: 23.2 -> 23.7 | Improved |
| lower threshold 0.7 versus 0.8 (ablation) | Diverse Learner · MS-COCO val (1% labeled) | mAP: 23.4 -> 23.7 | No clear effect |
| lower threshold 0.8 versus 0.7 without DL (ablation) | Ours (w/o DL) · MS-COCO val (1% labeled) | mAP: 22.6 -> 23.1 -> 22.7 | Mixed |
| add diverse learner to reduce teacher-student parameter similarity (ablation) | teacher-student pair · MS-COCO val (1% labeled) | layer-wise cosine similarity: 0.9987 -> 0.1498 | Improved |
| use Diverse Learner over Soft Teacher with 1% labels | Soft Teacher · MS-COCO val (1% labeled) | mAP: 21.62 -> 23.72 | Improved |
| use Diverse Learner over Soft Teacher with 5% labels | Soft Teacher · MS-COCO val (5% labeled) | mAP: 30.42 -> 31.92 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Instant Teaching — Instant-Teaching: An End-to-End Semi-Supervised Object Detection Framework (CVPR 2021) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unbiased Teacher for Semi-Supervised Object Detection (ICLR 2021 · similarity 0.93)
- Humble Teachers Teach Better Students for Semi-Supervised Object Detection (CVPR 2021 · similarity 0.93)
- Instant-Teaching: An End-to-End Semi-Supervised Object Detection Framework (CVPR 2021 · similarity 0.93)
- End-to-End Semi-Supervised Object Detection With Soft Teacher (ICCV 2021 · similarity 0.93)
- Unbiased Teacher v2: Semi-Supervised Object Detection for Anchor-Free and Anchor-Based Detectors (CVPR 2022 · similarity 0.91)
- Mind the Gap: Polishing Pseudo Labels for Accurate Semi-supervised Object Detection (AAAI 2023 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
