Research map / Detection and segmentation / CNN features
Gradient Harmonized Single-Stage Detector
Buyu Li, Yu Liu, Xiaogang Wang
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 cross-entropy with GHM-C for classification | RetinaNet-ResNet-50 · COCO minival | AP: 28.6 -> 35.8 | Improved |
| Replace Focal Loss with GHM-C for classification | RetinaNet-ResNet-50 · COCO minival | AP: 35.6 -> 35.8 | Improved |
| Use GHM-C + GHM-R instead of Focal Loss | RetinaNet-FPN-ResNeXt-101 · COCO test-dev | AP: 40.8 -> 41.6 | Improved |
| Add GHM-R for box regression instead of Smooth L1 | RetinaNet-ResNet-50 with GHM-C · COCO minival | AP: 35.8 -> 36.4 | Improved |
| Add GHM-R for box regression instead of Smooth L1 | RetinaNet-ResNet-50 with GHM-C · COCO minival | AP@IoU=0.5: 55.5 -> 54.6 | Hurt |
| Add GHM-R for box regression instead of Smooth L1 | RetinaNet-ResNet-50 with GHM-C · COCO minival | AP: 36.4 -> 37.4 | Improved |
| Replace Smooth L1 with Authentic Smooth L1 (ablation) | RetinaNet-ResNet-50 with GHM-C · COCO minival | AP: 35.8 -> 35.7 | No clear effect |
| Use 30 rather than 5 unit regions for GHM-C (ablation) | RetinaNet-ResNet-50 · COCO minival | AP: 33.4 -> 35.8 | Improved |
| Increase GHM-C unit regions from 30 to 40 (ablation) | RetinaNet-ResNet-50 · COCO minival | AP: 35.8 -> 35.4 | Hurt |
| Use unit-region approximation instead of the original gradient-density calculation (ablation) | GHM-C · COCO minival | average time per iteration: 13.675 -> 0.824 | Improved |
| Use unit-region approximation instead of the original gradient-density calculation (ablation) | GHM-C · COCO minival | AP: 35.9 -> 35.8 | No clear effect |
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.
- Faster R-CNN — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · 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
- The Devil Is in Gradient Entanglement: Energy-Aware Gradient Coordinator for Robust Generalized Category Discovery (CVPR 2026 · similarity 0.82)
- Spectral Gradient Descent Mitigates Anisotropy-Driven Misalignment: A Case Study in Phase Retrieval (ICML 2026 · similarity 0.81)
- Gradient-Weight Alignment as a Train-Time Proxy for Generalization in Classification Tasks (NeurIPS 2025 · similarity 0.81)
- Simplicity Bias via Global Convergence of Sharpness Minimization (ICML 2024 · similarity 0.81)
- Provably Reliable Classifier Guidance via Cross-Entropy Control (NeurIPS 2026 · similarity 0.81)
- Implicit Bias of Spectral Descent and Muon on Multiclass Separable Data (NeurIPS 2025 · similarity 0.81)
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.
