Strip R-CNN Framework for Elongated Object Detection

• Strip R-CNN is an object detection and segmentation framework that employs large strip convolutions to effectively capture the directional context of elongated objects.
• It uses sequential horizontal and vertical strip convolutions combined with specialized backbones and detection heads to enhance localization and angle regression.
• Quantitative evaluations on remote sensing and forensic datasets demonstrate improved mAP and parameter efficiency, validating its design over traditional square kernels.

Strip R-CNN is an object detection and instance segmentation framework characterized by the use of large strip convolutions. It is specifically tailored to improve detection performance for high aspect ratio, elongated objects commonly encountered in domains such as remote sensing and forensic image analysis. The Strip R-CNN family encompasses at least two distinct methodological lines: a convolution-centric architecture featuring StripNet backbones for large-scale object detection in aerial imagery (Yuan et al., 7 Jan 2025), and a pipeline for long, thin forensic trace segmentation merging PSPNet pre-segmentation and Mask R-CNN instance prediction (Zink et al., 2022).

1. Motivation and Theoretical Rationale

Standard convolutional architectures aggregate context isotropically via square kernels, which are suboptimal for slender, high aspect ratio objects (e.g., roads, runways, fibers). Square kernels include extraneous background and are less sensitive to directional dependencies, resulting in reduced localization accuracy—especially for orientation- and angle-sensitive tasks. Strip R-CNN remedies these deficiencies by employing sequential orthogonal strip convolutions, which focus contextual aggregation along one spatial axis at a time, enhancing representation for elongated structures (Yuan et al., 7 Jan 2025).

In forensic segmentation, Mask R-CNN struggles to segment long, non-axis-aligned traces due to fragmentation and orientation variance. Pre-segmenting with a semantic network (PSPNet) followed by geometrically guided mask merging addresses these weaknesses (Zink et al., 2022).

2. Mathematical Formulation of Strip Convolutions

The strip module operates as follows. Given a feature map $X \in \mathbb{R}^{C \times H \times W}$:

1. Local context capture: $Z = \mathrm{DWConv}_{5 \times 5}(X)$.
2. Sequential strip convolutions:
   • Horizontal, $Z_h = \mathrm{DWConv}_{k \times 1}(Z)$
   • Vertical, $Z_v = \mathrm{DWConv}_{1 \times k}(Z_h)$
3. Channel mixing: $Y = \mathrm{PWConv}_{1 \times 1}(Z_v)$
4. Input reweighting (attention-like): $\hat{Y} = X \odot Y$.

By selecting large $k$ (e.g., $k=19$), the receptive field approaches that of a $k \times k$ square kernel but with fewer parameters ($2kC$ vs. $k^2C$) (Yuan et al., 7 Jan 2025). Ablation studies on DOTA demonstrate that (19,19,19,19) sequential strip kernels yield the highest mAP (81.75%), outperforming square and dilated convolutions.

3. StripNet Backbone and Detection Heads

StripNet is constructed in two variants: StripNet-T and StripNet-S, both employing four stages with differing channel-depth configurations. Each stage comprises multiple "strip blocks"—each consisting of two residual sub-blocks:

• Strip block: Sequential strip convolutions with skip connections.
• Feed-Forward Network (FFN) block: Two 1×1 convolutions with GELU nonlinearity and skip connection.

Stages are separated by downsampling 3×3 convolutions with stride 2.

Detection heads are decoupled into specialized branches:

• Classification and Angle Head: Two fully connected layers (dimension 1024), shared.
• Localization Head: Separate, strengthened by strip convolutions; architecture is 3×3 conv → strip module → fully connected → regression for (x, y, w, h).
• Loss Function: Total loss $\mathcal{L} = L_c + L_l + L_\theta$, combining cross-entropy for classification, Smooth L1 for localization and angle regression (Yuan et al., 7 Jan 2025).

In the forensic segmentation pipeline (Zink et al., 2022), the backbone is ResNet-50+FPN enhanced with Swish activations for better thin-object mask quality, alongside a non-modified Mask R-CNN instance segmentation head.

4. Data Augmentation, Training, and Inference

Strip R-CNN models are typically pretrained on ImageNet and fine-tuned on task-specific datasets:

• Optimizer: AdamW ($\beta_1=0.9$, $\beta_2=0.999$, decay 0.05)
• Data Augmentation: Multi-scale training, cropping of overlapping $1024 \times 1024$ patches, random flips.
• Batch Size: 8 on 8×NVIDIA 3090 GPUs.
• StripNet-S: 30.5M parameters, 159 GFLOPs for $1024 \times 1024$ inputs.

In the segmentation variant, heavy regularization is employed:

• High dropout (0.4) after PSPNet encoder.
• Combined Dice + Binary Cross-Entropy loss ("Dice Entropy") for semantic mask training.
• Three-phase learning rate scheduling (warmup, plateau, decay).
• Advanced geometric and photometric augmentations.

Inference in segmentation proceeds by PSPNet pre-segmentation, Mask R-CNN instance prediction, and IoU-based mask merging with containment checks to yield final masks.

5. Quantitative Performance and Ablations

Remote Sensing Benchmarks (Yuan et al., 7 Jan 2025)

Dataset	mAP (Strip R-CNN-S)	Comparator/Gap	Model Size (M)
DOTA-v1.0	82.28% (single) / 82.75% (ensemble)	LSKNet-S: +2.57%	30.5
FAIR1M	48.26%	LSKNet-S: +0.39%	30.5
HRSC2016	98.70%	PKINet-S: +0.16%	30.5
DIOR-R	68.70%	PKINet-S: +1.67%	30.5

Ablations indicate:

• Kernel size (19): best for mAP.
• Sequential strip convolution improves over square/dilated/parallel strip designs by 0.20–0.30% mAP.
• Strip head design: adding strip convs to both (x, y) and (w, h) localization branches is optimal.

Forensic Segmentation (Zink et al., 2022)

Variant	Mean IoU
PSPNet, batch 8	0.810 ± 0.02
Swish vs. ReLU (deep layers)	0.829 ± 0.007
Mask R-CNN (vanilla) mAP@[.50:.95]	0.42
Strip R-CNN (PSPNet pre-seg + IoU merge) mAP@[.50:.95]	0.47 (+12%)

Qualitative improvements include continuous long-thin masks and reduction in false positive bubbles.

6. Strip R-CNN in Context: Architectural Innovations and Comparative Analysis

Strip R-CNN introduces directional, parameter-efficient long-range convolutional modules. Unlike large square-kernel methods, the sequential orthogonal strip design delivers focused context modeling—the $k \times 1$ and $1 \times k$ convolutions attend independently to horizontal and vertical aspects, beneficial for detecting roads, runways, rivers, and other slender features in aerial and forensic images (Yuan et al., 7 Jan 2025). Empirical correlation maps from the detection head demonstrate denser, longer-range dependencies over slender objects compared to conventional architectures.

The forensic variant (Zink et al., 2022) is not directly related architecturally but shares the motivation to enhance long object detection; here, semantic-then-instance two-stage prediction with IoU-based region merging is central.

7. Robustness, Generalization, and Implementation Considerations

Strip R-CNN models incorporate extensive regularization and augmentation for robust generalization. High encoder dropout, aggressive data augmentation (rotation, scaling, noise, blur, sharpening), and stratified dataset splitting reduce overfitting risk even in high intra-class variance scenarios (Zink et al., 2022). Optimized learning rate schedules guide convergence, and early stopping on held-out sets maintains generalization.

A plausible implication is that Strip R-CNN’s modular strip convolution approach can be extended to other domains featuring elongated or anisotropic object geometries. Its parameter efficiency supports practical deployment on larger imagery without prohibitive computational burden.

• "Strip R-CNN: Large Strip Convolution for Remote Sensing Object Detection" (Yuan et al., 7 Jan 2025)
• "Boosting Mask R-CNN Performance for Long, Thin Forensic Traces with Pre-Segmentation and IoU Region Merging" (Zink et al., 2022)