Research map / Detection and segmentation / Open-set and few-shot detection
Pursuing Better Decision Boundaries for Long-Tailed Object Detection via Category Information Amount
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes category information amount (volume of a class's perceptual manifold in detector embedding space) and an Information Amount-Guided Angular Margin (IGAM) loss that adjusts per-class decision spaces, evaluated on LVIS v1.0, COCO-LT and Pascal VOC object detection.
Problem
In object detection, category bias persists even when instance counts are balanced, so instance count alone does not explain per-class performance gaps; existing long-tail methods focus on instance counts and ignore intra-category complexity.
Approach
Define a category's information amount as the log-determinant volume of its covariance-based perceptual manifold, and use it to set class-specific angular margins in a cosine softmax loss so harder (higher-information) categories get larger decision spaces; a queue-based merged-covariance strategy makes the update cheap.
Key results
- On LVIS v1.0 with Faster R-CNN R-50-FPN, IGAM reaches 26.8 mAP and 19.0 AP_r versus 19.3 mAP and 1.1 AP_r for cross-entropy, with overall gains of 5.2-7.5 mAP across detectors.
- On COCO-LT, IGAM achieves 25.8 mAP (R-50) and 27.0 mAP (R-101), the best overall in the table.
- On Pascal VOC, IGAM reaches 77.7 and 78.6 mAP with R-50 and R-101 backbones, exceeding the second-best method by 0.8 and 1.1 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace cross-entropy with IGAM loss | Faster R-CNN, ResNet-50-FPN · LVIS v1.0 | mAP^b: 19.3 -> 26.8 | Helped |
| replace cross-entropy with IGAM loss | Faster R-CNN, ResNet-50-FPN · LVIS v1.0 rare categories | AP_r: 1.1 -> 19.0 | Helped |
| replace cross-entropy with IGAM loss | Faster R-CNN, ResNet-101-FPN · LVIS v1.0 | mAP^b: 20.9 -> 28.0 | Helped |
| replace cross-entropy with IGAM loss | Faster R-CNN, ResNet-50-FPN · COCO-LT | mAP^b: 24.5 -> 25.8 | Helped |
| replace cross-entropy with IGAM loss | Faster R-CNN, ResNet-101-FPN · COCO-LT | mAP^b: 26.0 -> 27.0 | Helped |
| replace cross-entropy with IGAM loss | Faster R-CNN, ResNet-50 · Pascal VOC 2007 test | mAP^b: 72.8 -> 77.7 | Helped |
| replace cross-entropy with IGAM loss | Faster R-CNN, ResNet-101 · Pascal VOC 2007 test | mAP^b: 73.5 -> 78.6 | Helped |
| set scale s to 20 instead of 30 (ablation) | IGAM, Faster R-CNN ResNet-50-FPN · LVIS v1.0 | mAP^b: 26.8 -> 26.2 | Hurt |
| set scale s to 10 instead of 30 (ablation) | IGAM, Faster R-CNN ResNet-50-FPN · LVIS v1.0 | mAP^b: 26.8 -> 16.8 | Hurt |
| set scale s to 40 instead of 30 (ablation) | IGAM, Faster R-CNN ResNet-50-FPN · LVIS v1.0 | mAP^b: 26.8 -> 25.9 | Hurt |
| set scale s to 50 instead of 30 (ablation) | IGAM, Faster R-CNN ResNet-50-FPN · LVIS v1.0 | mAP^b: 26.8 -> 24.6 | Hurt |
Surprising findings
- Category information amount correlates with per-class AP more strongly (Pearson about -0.66 to -0.70) than instance count does on relatively balanced Pascal VOC and long-tailed datasets.
- Cross-entropy baselines on LVIS achieve AP_r of only about 1-3 points, indicating near-failure on rare categories.
Limitations
- Performance on rare categories still lags frequent categories in absolute terms (AP_r 19-20 vs AP_f ~31-32 on LVIS).
- Hyperparameter s is sensitive (mAP drops from 26.8 at s=30 to 16.8 at s=10 on LVIS).
Experimental setup
Datasets: LVIS v1.0 COCO-LT Pascal VOC 2007 Pascal VOC 2012 MS COCO
Benchmarks: LVIS v1.0 COCO-LT Pascal VOC 2007 test
Metrics: mAP (bbox AP) AP_r AP_c AP_f AP_1-AP_4 (COCO-LT groups) per-class AP variance of class-wise AP as model bias
Baselines: Cross-Entropy Seesaw EQLv2 EFL C2AM RFS EQL DropLoss RIO BALMS BAGS LOCE
Builds on: Seesaw Loss EFL C2AM Ledoit-Péché nonlinear shrinkage
Closest papers by meaning
- C2AM Loss: Chasing a Better Decision Boundary for Long-Tail Object Detection (CVPR 2022 · similarity 0.91)
- Learning from Neighbors: Category Extrapolation for Long-Tail Learning (CVPR 2025 · similarity 0.86)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 2024 · similarity 0.86)
- SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection (CVPR 2025 · similarity 0.85)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.84)
- CQ-DINO: Mitigating Gradient Dilution via Category Queries for Vast Vocabulary Object Detection (NeurIPS 2025 · similarity 0.84)
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.