Research map / Detection and segmentation / CNN features
Learning 1-Bit Tiny Object Detector with Discriminative Feature Refinement
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 DFR-Det, a training-only discriminative feature refinement method that uses an information-bottleneck-motivated decoder with a ground-truth foreground mask to improve 1-bit tiny object detectors on aerial datasets.
Problem
1-bit detectors perform well on commonly sized objects but degrade severely on tiny objects, because high-level features from 1-bit convolutions poorly reveal discriminative foreground information.
Approach
Add a real-valued decoder that reconstructs the input image from FPN features P2-P4, with a loss that reconstructs foreground pixels (masked by ground-truth boxes) and suppresses background activations, motivated by an information bottleneck objective; the decoder is discarded at inference.
Key results
- 1-bit FCOS with DFR-Det reaches 12.8% AP on AI-TOD, versus 10.7% for IDa-Det and 13.5% for the real-valued FCOS.
- On AI-TOD with FCOS, DFR-Det improves AP by 4.6%, 3.8%, and 2.1% over BiDet, LWS-Det, and IDa-Det respectively.
- On DOTA-v2.0 with FCOS, DFR-Det reaches 30.9% AP, surpassing IDa-Det (28.7%) by 2.2 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DFR-Det to FCOS | 1-bit FCOS with ResNet-18 · AI-TOD test | AP: 10.7 -> 12.8 | Helped |
| add DFR-Det to RetinaNet | 1-bit RetinaNet with ResNet-18 · AI-TOD test | AP: 9.1 -> 10.6 | Helped |
| add DFR-Det to Faster-RCNN | 1-bit Faster-RCNN with ResNet-18 · AI-TOD test | AP: 7.8 -> 8.6 | Helped |
| add DFR-Det to FCOS | 1-bit FCOS with ResNet-18 · DOTA-v2.0 val | AP: 28.7 -> 30.9 | Helped |
| add DFR-Det to Faster-RCNN | 1-bit Faster-RCNN with ResNet-18 · DOTA-v2.0 val | AP: 29.1 -> 31.1 | Helped |
| add decoder using P2 features only (ablation) | 1-bit FCOS · AI-TOD | AP: 10.1 -> 10.9 | Helped |
| add decoder using P3 features only (ablation) | 1-bit FCOS · AI-TOD | AP: 10.1 -> 10.6 | Helped |
| add decoder using P4 features only (ablation) | 1-bit FCOS · AI-TOD | AP: 10.1 -> 10.2 | Helped |
| add decoder using P5 features only (ablation) | 1-bit FCOS · AI-TOD | AP: 10.1 -> 9.8 | Hurt |
| add decoder using P6 features only (ablation) | 1-bit FCOS · AI-TOD | AP: 10.1 -> 9.1 | Hurt |
| decode P2 and P3 features (ablation) | 1-bit FCOS · AI-TOD | AP: 10.1 -> 12.0 | Helped |
| decode P2, P3, and P5 features (ablation) | 1-bit FCOS · AI-TOD | AP: 10.1 -> 11.7 | Helped |
Surprising findings
- Reconstructing the full image (including background) gave only a 0.5% gain, while foreground-only reconstruction alone gave 0.5% and background suppression alone gave 1.6%, so the combined loss was needed for the 2.7% total gain.
- Decoding features from P5 and P6 degraded performance relative to the baseline when used alone.
Limitations
- Gains are shown mainly on aerial tiny-object datasets; generalization to other domains is not tested.
- Requires ground-truth boxes to build the foreground mask during training, limiting use to supervised settings.
- Hyperparameters (lambda, gamma) selected on AI-TOD with FCOS and reused elsewhere.
Experimental setup
Datasets: AI-TOD DOTA-v2.0 TinyPerson
Benchmarks: AI-TOD DOTA-v2.0
Metrics: AP AP0.5 AP0.75 APvt APt APs APm model size (MB) OPs (G)
Baselines: ReActNet BiDet LWS-Det IDa-Det LSQ Real-valued detector
Builds on: BiDet IDa-Det LWS-Det FCOS ReActNet
Closest papers by meaning
- Reviving the Discarded High-Resolution Feature for Transformer-Based Tiny Object Detection (NeurIPS 2026 · similarity 0.88)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.86)
- iDETR: Implicit DETR for Tiny Object Detection (NeurIPS 2026 · similarity 0.86)
- You Only Look One-Level Feature (CVPR 2021 · similarity 0.86)
- M2Det: A Single-Shot Object Detector Based on Multi-Level Feature Pyramid Network (AAAI 2019 · similarity 0.85)
- RemDet: Rethinking Efficient Model Design for UAV Object Detection (AAAI 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.