Research map / Detection and segmentation / CNN features
Uncertainty-Aware Gradient Stabilization for Small Object Detection
Huixin Sun, Yanjing Li, Linlin Yang, Xianbin Cao, Baochang Zhang
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| replace L2 with classification-based localization (L_CE) (ablation) | Faster R-CNN R-50 · VisDrone val (Faster R-CNN R-50, ablation) | AP: 21.3 -> 22.1 | Improved |
| add interval non-uniform (IN) quantized labels (ablation) | L_CE · VisDrone val (Faster R-CNN R-50, ablation) | AP: 22.1 -> 22.5 | Improved |
| add uncertainty minimization loss (lambda 0.5) (ablation) | L_CE+IN · VisDrone val (Faster R-CNN R-50, ablation) | AP: 22.5 -> 22.9 | Improved |
| increase UM weight lambda from 0.5 to 1.0 (ablation) | L_CE+IN+UM · VisDrone val (Faster R-CNN R-50, ablation) | AP: 22.9 -> 22.6 | Hurt |
| add uncertainty-guided refinement (UR) module (ablation) | L_CE+UM · VisDrone val (Faster R-CNN R-50, ablation) | AP: 22.9 -> 24.2 | Improved |
| raise UR perturbation rho from 0.5 to 1 (ablation) | UGS · VisDrone val (Faster R-CNN R-50, ablation) | AP: 24.2 -> 23.2 | Hurt |
| replace L2 with Smooth-L1 loss (ablation) | Faster R-CNN R-50 · VisDrone val (Faster R-CNN R-50, ablation) | AP: 21.3 -> 21.5 | No clear effect |
| replace L2 with gradient clipping (ablation) | Faster R-CNN R-50 · VisDrone val (Faster R-CNN R-50, ablation) | AP: 21.3 -> 21.6 | No clear effect |
| replace L2 with Bayesian YOLO (ablation) | Faster R-CNN R-50 · VisDrone val (Faster R-CNN R-50, ablation) | AP: 21.3 -> 22.0 | Improved |
| add UGS | Faster R-CNN · VisDrone val | AP: 21.3 -> 24.2 | Improved |
| add UGS | DINO-5scale · VisDrone val | AP: 35.5 -> 38.1 | Improved |
| add UGS | Rotated RetinaNet · SODA-A test | AP: 22.3 -> 26.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- Faster R-CNN — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · better on 4 of 4 shared results
- DQ-DETR — DQ-DETR: Dual Query Detection Transformer for Phrase Extraction and Grounding (AAAI 2023) · better on 1 of 1 shared result
- FCOS — FCOS: Fully Convolutional One-Stage Object Detection (ICCV 2019) · better on 1 of 1 shared result
- Cascade R-CNN — Cascade R-CNN: Delving Into High Quality Object Detection (CVPR 2018) · better on 1 of 1 shared result
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
- Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts (ICCV 2025 · similarity 0.87)
- Harnessing Uncertainty-aware Bounding Boxes for Unsupervised 3D Object Detection (ICCV 2025 · similarity 0.86)
- Rethinking Pseudo Labels for Semi-supervised Object Detection (AAAI 2022 · similarity 0.86)
- Focus, Align, and Sustain: Counteracting Gradient Dilution in Incremental Object Detection (ICML 2026 · similarity 0.86)
- Reconcile Prediction Consistency for Balanced Object Detection (ICCV 2021 · similarity 0.86)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.85)
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.
