Taming Dynamic Clutter: Variance-Driven Adaptive Gain Control for Bio-inspired Small Target Detection

Official implementation of the Adaptive Gain Control (AGC) module for bio-inspired Small Target Motion Detection (STMD).

Existing STMD models suffer from a false-alarm collapse under non-stationary environmental interference (e.g. rustling leaves, water glints, high-frequency flickers). We introduce a lightweight, plug-and-play adaptive shunting-inhibition operator that quantifies local temporal variance and maps it into a gain-control factor, suppressing heavy-tailed clutter while preserving genuine dim targets. The module integrates seamlessly into mainstream STMD architectures.

Method

Given a baseline STMD response Q, the AGC operator computes a variance-driven suppression factor omega and gates the response (Q̃ = omega · Q):

Step	Equation
Macro-compensated residual	E(x,y,t) = I(x,y,t) − I(x−φ, y−ψ, t−Δt) (Eq. 1)
Temporal variance (EMA)	σ²(t) = (1−γ)·σ²(t−Δt) + γ·E²(t) (Eq. 6)
Center-surround field	W_s = A[G_s2 − e·G_s3 − ρ]+ + B[G_s2 − e·G_s3 − ρ]− (Eq. 8)
Spatial clutter metric	C(x,y,t) = (σ² ∗ W_s)(x,y,t) (Eq. 7)
Suppression factor	omega(x,y,t) = exp(−C² / λ²) (Eq. 10)
Modulated output	Q̃ = omega · Q (Eq. 11)

Default parameters (paper Sec. 4.2): λ = 0.05, γ = 0.1, σ_s = 1.5.

The core operator lives in core/suppression.py (AdaptiveSuppression).

Repository structure

python/smalltargetmotiondetectors/
├── api/      High-level inference / evaluation entry points
├── core/     Neural-cascade operators (incl. suppression.py = AGC module)
├── model/    STMD baselines and their *WithSuppression (AGC) variants
├── util/     Kernels, I/O, evaluation metrics
├── demo/     Inference demos, dataset generation, and evaluation scripts
└── test/     Unit tests for the baselines
data/         Sample synthetic sequence + ground truth
demodata/     Sample image stream / videos for the inference demos

Models

Baselines and their AGC variants:

Baseline	AGC variant (Q̃ = omega · Q)
ESTMD	ESTMDWithSuppression
DSTMD	DSTMDWithSuppression
FeedbackSTMD	FeedbackSTMDWithSuppression
STFeedbackSTMD	STFeedbackSTMDWithSuppression

Installation

pip install -r requirements.txt

Requires Python ≥ 3.12. Dependencies: numpy, scipy, opencv-python, matplotlib, tqdm, torch.

Quick start

Run a model on the bundled image stream / video and visualize the output:

python start_by_python.py
# or directly:
python python/smalltargetmotiondetectors/demo/inference_image_stream.py
python python/smalltargetmotiondetectors/demo/inference_video.py

Programmatic use:

from smalltargetmotiondetectors.api import instancing_model, inference

model = instancing_model('STFeedbackSTMDWithSuppression')  # baseline + AGC
model.init_config()
result, run_time = inference(model, gray_frame)            # per-frame call

Reproducing the experiments

Datasets (download separately): Vision Egg (synthetic, generated locally), RIST, and the Infrared dim-small aircraft dataset. Update the dataset paths at the top of the relevant script before running.

Script (demo/)	Purpose
generate_vision_egg.py	Generate the synthetic Vision Egg sequences
add_flicker_to_real.py / add_target_to_real.py	Inject controlled flickers / targets into real video
multi_model_experiment.py	Synthetic comparison across 4 baselines × (with/without AGC)
standalone_evaluate_vision_egg.py	Vision Egg evaluation table
evaluate_real_world.py	RIST evaluation with injected flickers
sensitivity_analysis.py	Sensitivity to γ and λ
smoke_test_models.py	Quick end-to-end pipeline check

Metrics: AUC, Detection Rate, Precision, F1, False-Alarm Suppression Ratio (FASR).

Acknowledgements

This project builds on the open-source Small-Target-Motion-Detectors library by Mingshuo Xu, which provides the ESTMD/DSTMD/FeedbackSTMD/STFeedbackSTMD baselines. The STFeedbackSTMD baseline and its LPTC ego-motion compensation follow Wang et al., Bio-Inspired Small Target Motion Detection With Spatio-Temporal Feedback in Natural Scenes, IEEE TIP, 2024. Our contribution is the variance-driven AGC module and its integration with these baselines.

Citation

This work is under review. A citation entry will be added upon publication.

License

Released for research purposes. See the original Small-Target-Motion-Detectors repository for the baseline implementations' terms.