Dual-Statistic Synergy Gating (DSG) Module

• The DSG module is an adaptive channel-wise feature selection mechanism that integrates global means and peak-to-mean differences for effective gating in object detection.
• It employs the Dual-Statistic Synergy Operator (DSO) which fuses these statistics via a monotonic function followed by a lightweight 1×1 convolution and sigmoid activation.
• Empirical analyses show that integrating DSG in YOLO-DS enhances object discrimination efficiency, achieving higher accuracy with modest increases in computational cost.

The Dual-Statistic Synergy Gating (DSG) module is an adaptive channel-wise feature selection mechanism introduced in YOLO-DS, built specifically for explicit modeling of heterogeneous object responses within shared feature channels in one-stage object detectors. DSG is designed to enable “pick and choose” gating over concatenated multi-path bottleneck features (C2F), leveraging the joint modeling of global and local spatial channel statistics for improved object discrimination efficiency. Its construction is based on the Dual-Statistic Synergy Operator (DSO), which synthesizes channel-wise means and peak-to-mean differences into gating signals that are processed via a lightweight 1×1 convolution and sigmoid activation. Empirical results demonstrate that DSG outperforms existing gating approaches on the MS-COCO benchmark in terms of both accuracy and computational overhead (Huang et al., 26 Jan 2026).

1. Objective and Architectural Integration

The DSG module’s primary objective is fine-grained channel selection for feature tensors within the YOLO-DS framework. Within each C2F block—which concatenates outputs from several bottleneck paths (yielding a tensor of shape $B \times C' \times H \times W$)—DSG is positioned immediately after feature concatenation and before the subsequent convolutional layer. It operates on global pooled statistics (mean and peak-to-mean difference) computed from the pre-concatenation feature map ($B\times C\times H\times W$), and produces gating weights for the post-concat feature tensor ($B\times C'\times H\times W$).

2. Dual-Statistic Computation

DSG computes two key per-batch, per-channel summary statistics from the input feature map $x\in\mathbb{R}^{B\times C\times H\times W}$:

• Channel-wise mean: $$\mu_{b,c} = \frac{1}{HW}\sum_{h=1}^H\sum_{w=1}^W x_{b,c,h,w}$$, shape $B\times C\times 1\times 1$
• Channel-wise maximum: $m_{b,c} = \max_{h,w} x_{b,c,h,w}$, shape $B\times C\times 1\times 1$
• Peak-to-mean difference: $d_{b,c} = m_{b,c} - \mu_{b,c}$, shape $B\times C\times 1\times 1$

These statistics capture both global activation and localized channel sparsity, facilitating robust modeling of heterogeneous feature responses.

3. Dual-Statistic Synergy Operator (DSO)

The DSO fuses the computed mean and peak-to-mean difference into a single channel-wise decision response via the operator

$$\Phi(\mu, d) = (\mu + 1)(d + 1) - 1 = \mu d + \mu + d$$

where $y_{b,c} = \Phi(\mu_{b,c}, d_{b,c})$, shape $B\times C\times 1\times 1$. This operator is monotonic in both input statistics, ensuring that increases in either signal yield stronger downstream gating responses. Theoretical analysis confirms strict monotonicity ($\partial\Phi/\partial\mu > 0$, $\partial\Phi/\partial d > 0$).

4. Channel-Wise Gating Mechanism

Gating weights for the concatenated C2F tensor are produced as follows:

• Linear transformation: $z_{DSG} = W_{DSG} * y + b_{DSG}$, where $W_{DSG}\in\mathbb{R}^{C'\times C\times 1\times 1}$ and $b_{DSG}\in\mathbb{R}^{C'}$, yielding $z_{DSG}\in\mathbb{R}^{B\times C'\times 1\times 1}$
• Gate calculation: $w_{DSG} = \sigma(z_{DSG})$, shape $B\times C'\times 1\times 1$, where $\sigma$ is the sigmoid activation
• Feature reweighting: $x_{out} = w_{DSG} \odot x_{cat}$, shape $B\times C'\times H\times W$

Direct mapping from $C$ input channels to $C'$ concatenated channels is performed—DSG does not employ any intermediate reduction ratio.

Pseudocode Representation:

def DSG(x_in, x_cat): # x_in: B×C×H×W # x_cat: B×C'×H×W mu = x_in.mean(dim=[2,3], keepdim=True) m = x_in.amax(dim=[2,3], keepdim=True) d = m - mu y = (d + 1) * (mu + 1) - 1 z = conv1x1_DSG(y) w = torch.sigmoid(z) out = x_cat * w return out

5. Hyperparameters and Ablation Studies

Key hyperparameters for DSG are as follows:

• Direct $C\rightarrow C'$ mapping (no reduction ratio)
• Activation function: sigmoid
• $C'$ for C2F is defined as $(n+2)\cdot(C/2)$, where $n$ is the number of bottleneck paths

Ablation studies on YOLOv8-L (input size $640\times640$) report:

• “Mean” only variant: $\text{AP}+0.3\%$ (52.9→53.2)
• “Max” only variant: $\text{AP}-0.2\%$ decrease
• Full DSG with DSO: $\text{AP}+0.6\%$ (52.9→53.5), $+6.1$M params, $+4.9$ GFLOPs overhead

6. Quantitative Performance and Computational Overhead

Integration of DSG into the YOLOv8-L baseline yields:

• AP improvement: $+0.6\%$ (52.9→53.5)
• Model size: $43.7$M → $49.8$M parameters
• Computation: $165.2$ → $170.1$ GFLOPs
• Latency on T4 TensorRT: $9.06$ ms/img → $9.31$ ms/img ($+0.25$ ms)

Combined use with the complementary Multi-Path Segmented Gating (MSG) module results in a total AP lift of $+1.2\%$.

Variant	AP	Params (M)	GFLOPs	Latency (ms/img)
Baseline YOLOv8-L	52.9	43.7	165.2	9.06
DSG only	53.5	49.8	170.1	9.31
DSG+MSG	54.1	—	—	—

7. Visual and Qualitative Analysis

Eigen-CAM visualizations in (Huang et al., 26 Jan 2026) demonstrate that DSG enhances the compactness and object-centricity of activation maps. Small object regions exhibit highly localized peaks, while larger objects display suppressed uniform activations with focused responses on discriminative boundaries. This behavior is attributed to DSG’s fusion of global mean and local sparsity (peak-to-mean difference), providing scale-adaptive attention allocation. Qualitative analyses confirm superior feature discrimination for varied object scales compared to baseline channel-attention mechanisms.

---

The DSG module operationalizes a dual-statistic synergy-based gating mechanism, synthesizing global and local channel statistics into effective feature selection weights. Only the complete DSO yields significant accuracy improvements with minimal added overhead, and qualitative analysis substantiates its role in efficient heterogeneous object response modeling (Huang et al., 26 Jan 2026).