Research map / Detection and segmentation / Open-set and few-shot detection
Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction
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 2DRCL, a detection pre-training framework combining holistic-local contrastive learning, instance-aware dynamic resampling, and dual (appearance and semantic) reconstruction to improve tail-class object detection on COCO and LVIS v1.0.
Problem
Pre-training for object detection with randomly or ImageNet-initialized detection heads performs poorly on long-tailed distributions due to extreme data imbalance and simplicity bias, especially for tail classes.
Approach
Combines MoCo-style holistic contrastive learning with RoI-level local contrastive learning on class-agnostic proposals, a dynamic image- and instance-level resampling whose weighting shifts over pre-training epochs, and an autoencoder-based appearance reconstruction plus masked semantic feature consistency loss to counter simplicity bias.
Key results
- On COCO with Mask R-CNN R50-FPN, 2DRCL reaches 41.4 AP^bb versus 39.4 for AlignDet and 38.3 for the ImageNet supervised baseline.
- On LVIS v1.0 with Mask R-CNN R50-FPN at 1x, 2DRCL reaches 23.9 AP and 11.9 AP_r versus 22.6 AP and 10.3 AP_r for AlignDet.
- Combined with ECM on LVIS v1.0 (2x, Mask R-CNN R50-FPN), the method reports 28.3 AP^bb and 20.4 AP_r.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HCL (holistic contrastive learning) to the baseline (ablation) | Mask R-CNN R50 pre-training · LVIS v1.0 | AP_r^bb: 9.1 to 10.5 | Helped |
| add dynamic rebalancing (DRB) on top of HLCL (ablation) | 2DRCL pre-training · LVIS v1.0 | AP_r^bb: 10.8 to 14.3 | Helped |
| add Dual Reconstruction (AR and SR) on top of HLCL+DRB (ablation) | 2DRCL pre-training · LVIS v1.0 | AP_r^bb: 14.3 to 15.2 | Helped |
| add LCL (local contrastive learning) alone (ablation) | Mask R-CNN R50 pre-training · LVIS v1.0 | AP^bb: 22.7 to 21.9 | Hurt |
| add 2DRCL pre-training before existing long-tailed fine-tuning methods | RFS, IRFS, EQLv2, SeeSaw, ECM, ROG, LOCE, BACL · LVIS v1.0 (R50-FPN, 12 epochs or 24+6/12+12) | AP^bb: Gains of 0.2 to 1.3 AP^bb across methods (e.g., RFS 22.7 to 23.9, ECM 26.5 to 27.3) | Helped |
| add 2DRCL pre-training before existing long-tailed fine-tuning methods | RFS, EQLv2, SeeSaw, ECM, BACL · LVIS v1.0 rare-class AP (AP_r^bb) | AP_r^bb: Gains range from about +0.1 to +2.0 (e.g., SeeSaw 14.7 to 17.5, ECM 17.0 to 19.2) | Helped |
| add 2DRCL pre-training with same total epochs (6 pretrain + 6 finetune) before RFS, EQL, Seesaw, ECM | Long-tailed fine-tuning methods · LVIS v1.0 same-epoch comparison | AP^bb: Average +0.5 AP^bb and +1.3 AP_r reported in text | Helped |
| 2DRCL pre-training vs. ImageNet-supervised init (AlignDet comparison) on COCO | Mask R-CNN R50-FPN · COCO val | AP^bb: 41.4 vs 38.3 supervised and 39.4 AlignDet | Helped |
| add 2DRCL pre-training on ATSS one-stage detector | ATSS R50 · LVIS v1.0 2x | AP_r^bb: 17.7 vs 16.6 (ECM) and 14.5 (Focal Loss) | Helped |
| add 2DRCL pre-training on frequent-class AP for ECM-style fine-tuning | Frequent classes · LVIS v1.0 Mask R-CNN R50 (ablation/per-group) | AP_f^bb: Some gains (e.g., ECM 31.7 to 32.5 in one row) but ATSS frequent-class AP falls below Focal Loss (30.8 vs 31.8) | Mixed |
Limitations
- Relatively high computational cost, acknowledged by the authors.
- Some feature-level advantages are diminished during downstream fine-tuning, according to the bias analysis.
- Slight accuracy decrease on common and frequent classes in some settings (e.g., ATSS frequent-class AP and ECM+Ours on ResNet-101).
Experimental setup
Datasets: COCO LVIS v1.0 COCO-LT
Benchmarks: COCO LVIS v1.0
Metrics: AP AP50 AP75 AP_r AP_c AP_f
Baselines: MoCo v2 SimCLR BYOL SoCo AlignDet DenseCL Self-EMD SlotCon EQLv2 Seesaw ECM LOCE
Builds on: MoCo AlignDet Repeat Factor Sampling InfoNCE
Closest papers by meaning
- Balanced Contrastive Learning for Long-Tailed Visual Recognition (CVPR 2022 · similarity 0.90)
- Equalization Loss v2: A New Gradient Balance Approach for Long-Tailed Object Detection (CVPR 2021 · similarity 0.89)
- Decoupled Contrastive Learning for Long-Tailed Recognition (AAAI 2024 · similarity 0.89)
- SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection (CVPR 2025 · similarity 0.88)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.88)
- OFBD: Object-Focused Background Debiasing for Long-Tailed Learning (NeurIPS 2026 · similarity 0.87)
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.