Research map / Detection and segmentation / CNN features
SET: Spectral Enhancement for Tiny Object Detection
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 SET, a plug-in training-time method for tiny object detection that combines hierarchical background smoothing in FPN features (suppressing high-frequency background noise) with adversarial feature perturbation injection, motivated by a frequency-domain occlusion analysis.
Problem
Tiny object detection is hard because tiny objects lose distinct features after encoding and are buried in noisy high-frequency background clutter, and generic detectors perform poorly on them.
Approach
Uses a frequency-domain occlusion analysis to show tiny objects benefit from removing high-frequency information; the HBS module decouples foreground/background via a GT mask and smooths background features with a channel-reduce-then-expand conv block, while the API module injects L2 feature-level adversarial perturbations during training only, adding no inference cost.
Key results
- On AI-TOD, SET lifts RFLA from 21.7% to 24.9% AP (+3.2%) and DINO-5scale from 23.2% to 26.6% AP.
- On AI-TOD, FCOS with SET improves from 12.0% to 14.2% AP, and FCOS with P2-P6 features improves from 15.1% to 18.0% AP.
- On MS COCO, FCOS with SET reaches 37.4 AP vs 36.4 baseline (+1.0), with APt rising 19.6 to 23.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SET to RFLA | RFLA · AI-TOD test | AP: 21.7 -> 24.9 | Helped |
| Add SET to DINO-5scale | DINO-5scale · AI-TOD test | AP: 23.2 -> 26.6 | Helped |
| Add SET to RFLA | RFLA · VisDrone2019 val | AP: 27.2 -> 28.5 | Helped |
| Add SET to FCOS | FCOS · DOTA-v2.0 val | AP: 31.8 -> 34.7 | Helped |
| Add SET to AutoAssign | AutoAssign · DOTA-v2.0 val | AP: 33.8 -> 35.2 | Helped |
| Add SET to FCOS | FCOS · COCO val2017 | APt: 19.6 -> 23.3 | Helped |
| Add adaptive smoothing to full features without foreground-background decoupling (ablation) | FCOS · AI-TOD test | AP: 12.0 -> 12.7 | Helped |
| Add HBS with foreground-background decoupling (ablation) | FCOS · AI-TOD test | AP: 12.0 -> 13.9 | Helped |
| Add API without HBS (ablation) | FCOS · AI-TOD test | AP: 12.0 -> 13.2 | Helped |
| Use both HBS and API (ablation) | FCOS · AI-TOD test | AP: 12.0 -> 14.2 | Helped |
| Set HBS kernel-size function g(Si) to 7 (ablation) | FCOS with HBS · AI-TOD test | AP: 13.9 -> 12.1 | Hurt |
Surprising findings
- Static high-frequency removal of full features is reported to help very tiny and tiny objects while hurting larger objects.
- The QueryDet small-object detection method decreases RetinaNet baseline performance on AI-TOD.
Limitations
- Gains on very tiny objects remain small in absolute terms (e.g., APvt around 2-13% on AI-TOD).
- Adversarial training increases training time noticeably.
- Hyperparameters (r, rho, lambda, kernel size rule) are tuned on AI-TOD and may not transfer uniformly.
Experimental setup
Datasets: AI-TOD VisDrone2019 DOTA-v2.0 COCO 2017
Benchmarks: AI-TOD VisDrone2019 DOTA-v2.0 COCO 2017
Metrics: AP AP0.5 AP0.75 APvt APt APs APm
Baselines: RFLA SR-TOD DNTR DotD NWD (DetectoRS w/NWD) QueryDet DCFL FCOS RetinaNet DINO-5scale
Builds on: FCOS adversarial training for feature perturbations Feature Pyramid Network RFLA DINO
Closest papers by meaning
- Expanding Low-Density Latent Regions for Open-Set Object Detection (CVPR 2022 · similarity 0.85)
- Multi-Spectral Salient Object Detection by Adversarial Domain Adaptation (AAAI 2020 · similarity 0.84)
- Multi-Perspective Data Augmentation for Few-shot Object Detection (ICLR 2025 · similarity 0.83)
- Learning 1-Bit Tiny Object Detector with Discriminative Feature Refinement (ICML 2024 · similarity 0.83)
- Unsupervised Object Detection with Theoretical Guarantees (NeurIPS 2024 · similarity 0.82)
- SNIDA: Unlocking Few-Shot Object Detection with Non-linear Semantic Decoupling Augmentation (CVPR 2024 · similarity 0.82)
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.