Research map / Detection and segmentation / CNN features
Feature Information Driven Position Gaussian Distribution Estimation 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 a plug-and-play module for FPN-based detectors that enhances weak tiny-object features using an unsupervised information-entropy-derived information map and a supervised Gaussian-mixture position distribution map.
Problem
Generic detectors degrade sharply on tiny objects because downsampling weakens features of objects covering very few pixels; prior attention and fusion methods rely on unreliable heuristic maps.
Approach
Estimate per-pixel information amount of the P2 feature by minimizing an entropy-coding (Gaussian-scale) loss to get an information map, and predict a size-aware position Gaussian mixture distribution map from bounding boxes with multi-scale features; both modulate P2 to enhance tiny-object regions.
Key results
- On VisDrone2019, plugging the method into DetectoRS raises AP from 26.3 to 28.3 and APt from 7.5 to 12.6 (ablation table, same setting).
- Combined with RFLA on VisDrone2019, the method reports the best APvt of 7.4 among compared methods.
- On AI-TOD, DetectoRS with the method reaches AP 28.3 (+2.0) and AP0.5 48.5 (+4.6) over the DetectoRS baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the proposed feature enhancement to RFLA | RFLA · VisDrone2019 | AP: 27.2 -> 29.0 | Helped |
| Add the proposed feature enhancement to DetectoRS | DetectoRS · AI-TOD | AP: 14.6 -> 24.3 | Helped |
| Add the proposed feature enhancement to DetectoRS | DetectoRS · AI-TODv2 | AP: 16.1 -> 25.5 | Helped |
| Add Pixels Feature Information Modeling (ablation) | DetectoRS · VisDrone2019 | AP: 26.3 -> 28.2 | Helped |
| Add Position Gaussian Distribution Prediction (ablation) | DetectoRS · VisDrone2019 | AP: 26.3 -> 27.6 | Helped |
| Replace instance-size-dependent Gaussian scaling with fixed α=1 (ablation) | DetectoRS w/ ours · VisDrone2019 | AP: 28.3 -> 27.9 | Hurt |
| Replace the smooth position distribution map with a binary bounding-box mask (ablation) | DetectoRS w/ ours · VisDrone2019 | AP: 28.3 -> 27.9 | Hurt |
| Replace the position Gaussian distribution map with a self-attention weight map (ablation) | DetectoRS w/ ours · VisDrone2019 | AP: 28.3 -> 27.7 | Hurt |
| Multiply prediction inputs by σ instead of using the proposed guidance (ablation) | DetectoRS w/ ours · VisDrone2019 | AP: 28.3 -> 28.3 | No clear effect |
| Multiply prediction inputs by (1 + σ) (ablation) | DetectoRS w/ ours · VisDrone2019 | AP: 28.3 -> 28.1 | Hurt |
| Concatenate σ with prediction input features (ablation) | DetectoRS w/ ours · VisDrone2019 | AP: 28.3 -> 27.9 | Hurt |
| Fuse the two enhanced features by element-wise multiplication (ablation) | DetectoRS w/ ours · VisDrone2019 | AP: 28.3 -> 27.5 | Hurt |
Surprising findings
- Multiplying the information map by sigma (instead of 1+sigma) gives the best APs but large drops on other metrics, attributed to loss of context where sigma is near zero.
Limitations
- Gains on APvt remain low in absolute terms (e.g., 3.5 for DetectoRS on VisDrone2019), so very tiny objects remain hard.
- Relies on hand-set scaling factors per object size class (4, 6, 8, 10) and threshold-based post-processing for the Gaussian target map.
- Ablations and analysis are limited to VisDrone2019 with DetectoRS.
Experimental setup
Datasets: VisDrone2019 AI-TOD AI-TODv2
Benchmarks: VisDrone2019 AI-TOD AI-TODv2
Metrics: AP AP0.5 AP0.75 APvt APt APs
Baselines: Faster R-CNN Cascade R-CNN DetectoRS NWD-RKA RFLA CEASC PKS R-CNN Salience DETR SR-TOD
Builds on: Feature Pyramid Networks DetectoRS Generalized Divisive Normalization CBAM SR-TOD
Closest papers by meaning
- Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection (NeurIPS 2020 · similarity 0.84)
- Unsupervised Object Detection with Theoretical Guarantees (NeurIPS 2024 · similarity 0.83)
- Gaussian-Det: Learning Closed-Surface Gaussians for 3D Object Detection (ICLR 2025 · similarity 0.83)
- Self-Supervised Learning of Intertwined Content and Positional Features for Object Detection (ICML 2025 · similarity 0.83)
- Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances (CVPR 2025 · similarity 0.82)
- Leveraging Imagery Data with Spatial Point Prior for Weakly Semi-supervised 3D Object Detection (AAAI 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.