Research map / Detection and segmentation / CNN features
HS-FPN: High Frequency and Spatial Perception FPN 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 HS-FPN, an FPN variant with a high-frequency perception module (DCT high-pass filtered responses used as channel and spatial masks) and a pixel-level spatial dependency perception module, improving tiny object detection on AI-TOD and a DOTA subset.
Problem
Feature Pyramid Networks do not directly enhance features of tiny objects, which occupy very few feature-map pixels, and they lack spatial perception to counter misalignment from upsampling, hurting tiny object detection.
Approach
Use a DCT-based high-pass filter to produce high-frequency responses that serve as channel and spatial attention masks (HFP) on each lateral connection, plus a block-wise pixel-level cross-attention between adjacent pyramid levels (SDP) to capture spatial dependencies missing in FPN.
Key results
- Cascade R-CNN with HS-FPN reaches 23.6 AP on AI-TOD vs 20.2 with FPN (+3.4).
- Faster R-CNN with HS-FPN reaches 20.3 AP on AI-TOD vs 18.3 with FPN (+2.0).
- On DOTA_mini10, Cascade R-CNN + HS-FPN achieves 50.9 AP vs 49.4 with FPN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FPN with HS-FPN | Faster R-CNN with ResNet-50 · AI-TOD test | AP: 18.3 -> 20.3 | Helped |
| Replace FPN with HS-FPN | Cascade R-CNN with ResNet-50 · AI-TOD test | AP: 20.2 -> 23.6 | Helped |
| Replace FPN with HS-FPN | Cascade R-CNN · DOTA_mini10 | AP: 49.4 -> 50.9 | Helped |
| Add channel path only (ablation) | FPN in Cascade R-CNN with ResNet-50 · AI-TOD test | AP: 20.2 -> 21.6 | Helped |
| Add spatial path only (ablation) | FPN in Cascade R-CNN with ResNet-50 · AI-TOD test | AP: 20.2 -> 21.6 | Helped |
| Add HFP with both channel and spatial paths (ablation) | FPN in Cascade R-CNN with ResNet-50 · AI-TOD test | AP: 20.2 -> 22.4 | Helped |
| Add SDP without HFP (ablation) | FPN in Cascade R-CNN with ResNet-50 · AI-TOD test | AP: 20.2 -> 21.3 | Helped |
| Add both HFP and SDP (ablation) | FPN in Cascade R-CNN with ResNet-50 · AI-TOD test | AP: 20.2 -> 23.6 | Helped |
| Set high-pass filter α to 0.25 instead of 0.00 (ablation) | HS-FPN · AI-TOD test | AP: 21.7 -> 22.4 | Helped |
| Increase α beyond 0.25 (ablation) | HS-FPN · AI-TOD test | AP: 22.4 -> 22.1; 22.4 -> 21.9; 22.4 -> 21.5 | Mixed |
Surprising findings
- Filtering out low-frequency components is only beneficial within a moderate range; alpha=0.25 is best and alpha=1 (blocking all frequencies) degrades performance below the FPN baseline's neighbourhood.
Limitations
- Added FLOPs and parameters relative to FPN, with SDP cost growing with patch size.
- Anchor base size set to 2 for anchor-based models, a tuned setting specific to tiny objects.
- Evaluated mainly on aerial datasets (AI-TOD and a self-built DOTA subset).
Experimental setup
Datasets: AI-TOD DOTA DOTA_mini10
Benchmarks: AI-TOD DOTA_mini10
Metrics: AP AP50 AP75 AP_vt AP_t AP_s AP_m FLOPs Params
Baselines: FPN ATSS MENet RetinaNet Cascade RPN Grid R-CNN Faster R-CNN NWD RFLA DetectoRS Cascade R-CNN
Builds on: Feature Pyramid Network FCANet GFNet Vision Transformer
Closest papers by meaning
- NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection (CVPR 2019 · similarity 0.88)
- Feature Pyramid Networks for Object Detection (CVPR 2017 · similarity 0.88)
- AugFPN: Improving Multi-Scale Feature Learning for Object Detection (CVPR 2020 · similarity 0.87)
- M2Det: A Single-Shot Object Detector Based on Multi-Level Feature Pyramid Network (AAAI 2019 · similarity 0.86)
- FD2-Net: Frequency-Driven Feature Decomposition Network for Infrared-Visible Object Detection (AAAI 2025 · similarity 0.86)
- FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network (NeurIPS 2025 · similarity 0.83)
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.