PLUSNet: Modular Small Object Detection

• PLUSNet is a detection framework that integrates three key modules to address small object detection challenges across feature extraction, labeling, and detection head design.
• The Hierarchical Feature Purifier (HFP) uses FFT-based high-pass filtering to suppress low-frequency noise and enhance crucial high-frequency details for better localization.
• The Multiple Criteria Label Assignment (MCLA) and Frequency Decoupled Head (FDHead) work together to boost positive sample rates and decouple semantic from contour features, leading to significant mAP improvements.

PLUSNet is a high-quality detection framework designed to address critical bottlenecks in small object detection pipelines by holistically optimizing upstream feature purification, midstream label assignment, and downstream detection head specialization. The architecture introduces three sequential and modular components—Hierarchical Feature Purifier (HFP), Multiple Criteria Label Assignment (MCLA), and Frequency Decoupled Head (FDHead)—that collectively improve small-object detection performance across diverse multi-scale scenarios and integrate readily into existing object detectors (Wang et al., 29 Apr 2025).

1. Motivation and Architectural Overview

The canonical small object detection pipeline consists of three stages: upstream feature extraction and fusion (typically via Feature Pyramid Networks, FPNs), midstream label assignment for proposals or anchors, and downstream detection heads for classification and regression. PLUSNet identifies three primary sources of small-object detection failure: noisy, low-frequency–dominated features in the bottom FPN levels; label assigners relying solely on the Intersection over Union (IoU) criterion, resulting in sparse positives for small/tiny objects; and detection heads using shared features for both classification and regression, thus failing to decouple context and detail requirements.

To address these challenges in a unified manner, PLUSNet inserts three sequential modules:

Module	Position in Pipeline	Primary Role
Hierarchical Feature Purifier (HFP)	Upstream (FPN)	Suppress low-frequency noise, enhance detail
Multiple Criteria Label Assignment (MCLA)	Midstream (Labeling)	Increase small-object positive samples
Frequency Decoupled Head (FDHead)	Downstream (Detection Head)	Specialize semantic/contour sub-branches

This modularity allows simple plug-and-play integration with one-stage, two-stage, anchor-free, and query-based detectors without major architectural modifications.

2. Hierarchical Feature Purifier (HFP)

HFP operates at low-level FPN layers, targeting the suppression of overwhelming low-frequency (“semantic”) noise and the enhancement of high-frequency details critical for small-object representation. Given an FPN feature map $X_l \in \mathbb{R}^{H \times W \times C}$ at level $l$ (with relay level $r$ controlling purification depth):

1. Compute frequency transform: $S = \mathrm{FFT}(X_l)$.
2. Apply an adaptive high-pass mask $M \in \{0,1\}^{H\times W}$ with kernel size $K = \mu \times \frac{r - l}{r}$, where $\mu$ controls filter strength.
3. Blur low-frequency regions (central square of size $K \times K$ set to zero) and reconstruct:

$$\tilde{X}_l = \mathrm{IFFT}(M \odot \mathrm{FFT}(X_l)) \times \omega + X_l$$

with residual-fusion weight $\omega$.

As a result, HFP endows the detector with cleaner, detail-rich multi-scale features for improved localization and recognition of small objects.

3. Multiple Criteria Label Assignment (MCLA)

Traditional IoU-only assigners are biased against small objects, yielding too few positive samples. MCLA addresses this by fusing three complementary criteria:

• IoU score, $S_1$
• Position Offset Consistency (POC), $S_2 = (1 + \sqrt{20 \cdot E_1^{norm}})^{-1}$, where $E_1^{norm}$ is a normalized position offset
• Shape-Size Consistency (SCC), $S_3 = (1 + \sqrt{0.25 \cdot E_2})^{-1}$, with $E_2$ as the squared error in width/height

The final assignment score:

$$S^{ij} = \frac{\lambda_1 S_1^{ij} + \lambda_2 S_2^{ij} + \lambda_3 S_3^{ij}}{\lambda_1 + \lambda_2 + \lambda_3}$$

Typical hyperparameters are $\lambda_1 = 1.0$ (IoU), $\lambda_2 = 3.0$ (POC), and $\lambda_3 = 1.0$ (SCC). Top-$K$ or thresholded proposals are then assigned as positives per ground-truth object. Empirical simulation demonstrates that MCLA substantially increases the positive rate among very small objects compared to MaxIoU assigners.

4. Frequency Decoupled Head (FDHead)

FDHead decouples RoI features in the frequency domain, allowing the classification branch to focus on low-frequency semantic attributes while dedicating the regression branch to high-frequency contour information. Specifically, for an input RoI feature $F_{roi}$:

• Low-pass branch: $$F_L = \mathrm{BR}_1(\mathrm{LowPass}_{D_l}(\mathrm{FFT}(F_{roi})))$$, used for classification through two fully connected layers
• High-pass branch: $$F_H = \mathrm{BR}_2(\mathrm{HighPass}_{D_h}(\mathrm{FFT}(F_{roi})))$$, used for regression via convolutional blocks and global average pooling

Separate linear layers yield the final logits and bounding-box offsets; losses remain cross-entropy for classification and Smooth-L1 for regression.

5. Training Protocols and Hyperparameter Settings

PLUSNet is trained using standard schedules (e.g., 12 epochs at learning rate 0.01, decayed at epochs 8 and 11), SGD with momentum 0.9, weight decay $1 \times 10^{-4}$, and batch sizes tailored to the dataset. Key ablation-identified hyperparameters include:

• HFP: $\mu=0.05$, $\omega=0.3$, relay level $r=2$
• MCLA: $\lambda_1=1.0$, $\lambda_2=3.0$, $\lambda_3=1.0$
• FDHead: $D_l{\sim}0.85$, $D_h{\sim}0.10$

No architectural changes are required for inference, as MCLA only affects training. HFP replaces standard FPN upsample/add at select levels, FDHead replaces the conventional detection head.

6. Experimental Validation and Observed Limitations

Extensive experiments on SODA-D (autonomous driving), AI-TOD (aerial/drone), and MS COCO benchmarks demonstrate that PLUSNet achieves significant and consistent improvements. Representative results include:

Dataset / Baseline	mAP Baseline	mAP PLUSNet	Notable Small Obj Gains
SODA-D / R50 FR-CNN	28.9	32.0	AP_eS: 13.8→15.4
AI-TOD / R50 FR-CNN	11.1	21.7	AP_vt: 0→8.3
COCO mini-val / R50 FR-CNN	37.4	39.4	AP_s: +2.0

Simulation with 2000 synthetic GT boxes shows that, for objects ≤12px, standard MaxIoU assigns 0% positives, while MCLA assigns 7.4%. Qualitative studies confirm that PLUSNet recovers small targets (traffic lights, signs, distant pedestrians, ships, windmills) missed by baselines.

The addition of FFT/IFFT-based modules incurs a computational overhead: GFLOPs increase from ~207 to ~481, and parameter count from ~41M to ~47.6M. This overhead may constrain real-time or mobile applications, suggesting future research into lightweight or learned frequency transforms.

7. Contributions, Key Insights, and Prospective Directions

PLUSNet’s main contributions are:

• Holistic analysis identifying core pipeline bottlenecks in small-object detection
• Three plug-and-play modules—HFP, MCLA, FDHead—targeting separate pipeline weaknesses
• Strong performance across baseline architectures on diverse benchmarks, with gains up to +10 AP on aerial data

Key findings establish that frequency-domain purification and head-branch decoupling are highly effective for small-object recall and precision, and that multi-criterion assignment circumvents IoU’s inherent bias against tiny objects. The design allows seamless integration into prevalent detector frameworks.

Identified limitations include increased computational cost and the need for more efficient frequency operations. Future research directions involve replacing FFT/IFFT with lightweight or learnable alternatives, jointly optimizing all pipeline stages end-to-end, and extending this paradigm to domain adaptation, video detection, and ultra-small object (<8px) scenarios (Wang et al., 29 Apr 2025).