PaperlayerView exact revision on arXiv

Source-linked AI summary

ScopeMamba-YOLO: Widening the Perceptual Scope Inward and Outward for Small Object Detection in Remote Sensing Imagery

Junjie Fan, Yijun Mai, Linduo Wei, Jiayu Rao, Junmin Bao, Qiushi Jin, Guijia Li, Yong Qi

arXiv:2609.10156v1cs.CV

TL;DR

Small-object detection in aerial imagery must retain weak high-resolution evidence while recovering broad contextual support. ScopeMamba-YOLO addresses this with off-path, zero-gated selective scanning alongside anisotropic extraction and scale-adaptive regression, improving reported detection performance across VisDrone-2019 and AI-TOD while retaining important computational and generalization boundaries.

  • Problem

    Small-object detection must preserve fine-grained evidence while maintaining broad contextual support, but direct main-path selective scanning can interfere with the detail-carrying representation.

  • Method

    ScopeMamba-YOLO decouples contextual modeling from the convolutional stream through off-path, zero-gated selective scanning in CGCM and SS-PAN, with AMS and SA-DFL supporting efficient multiscale detection.

  • Results

    The proposed design favors selective scanning over tested alternatives, with selective scanning reaching 0.5078 mAP50 at the neck versus 0.5013 for EMA and SA-DFL outperforming tested quality-branch variants.

  • Takeaways & Limitations

    Off-path context modeling, scale-adaptive regression, and anisotropic extraction are favored over direct main-path insertion, quality branches, loss shaping, and added depth under the tested configurations.

  • Takeaways & Limitations

    The evaluation is limited to 640×640 inputs without pretraining and to YOLO-style detection, while stride-4 and native-resolution context pathways retain non-negligible computation.

Abstract

from arXiv · show

Small object detection in unmanned aerial vehicle (UAV) and remote sensing imagery requires preserving high-resolution detail while modeling long-range context. Adding a stride-4 detection level and removing the stride-32 stage benefits tiny targets but weakens peripheral spatial support, whereas directly inserting selective scanning into the main feature path can interfere with weak local cues. We propose ScopeMamba-YOLO, built around an off-path, zero-gated selective-scanning principle that decouples contextual modeling from the convolutional stream. The principle is instantiated by a Cascaded Global-Context Module (CGCM) in the backbone and a Selective-Scan PAN (SS-PAN) in the neck. An Adaptive Multi-scale Strip (AMS) Block reduces the cost of high-resolution feature extraction, while a Scale-Adaptive DFL (SA-DFL) head reallocates distributional support and regression capacity across scales with only 0.008M additional parameters. Controlled experiments show that matched main-path selective scanning reduces mAP50 by 0.98 pp, whereas off-path CGCM improves the final configuration by 0.67 pp over the three-seed no-CGCM mean; operator controls indicate that this gain is not explained by auxiliary branch capacity alone. ERF analysis further shows that the complete context pathway increases the peripheral energy ratio from 0.008 to 0.090 at stride 8. On VisDrone-2019, ScopeMamba-S achieves 50.8% mAP50 with 3.57M parameters, exceeding YOLOv8s by 10.8 pp while using 32% of its parameters; ScopeMamba-M reaches 52.6% mAP50 with 6.48M parameters. Consistent improvements are also observed on AI-TOD, especially for very-tiny and tiny objects.

I. INTRODUCTION

Small-object detection in aerial imagery must preserve weak high-resolution cues while incorporating broad scene context. ScopeMamba-YOLO addresses this tension by reallocating resolution, using anisotropic extraction, and decoupling selective scanning from the convolutional stream.

  • I. INTRODUCTION: Aerial small objects provide sparse, easily attenuated evidence yet require local detail and long-range context to distinguish targets from clutter.The paper frames this as widening perceptual scope inward toward object detail and outward toward scene dependencies.
  • I. INTRODUCTION: Adding a stride-4 detection level while removing stride-32 improves small-object accuracy but weakens the deepest contextual support.This resolution–context trade-off motivates the proposed architecture.
  • I. INTRODUCTION: Matched main-path scanning reduces mAP50 by 0.98 pp, whereas off-path CGCM improves mAP50 by 0.67 pp over the no-CGCM control.These observations motivate using selective scanning as complementary contextual modulation rather than a replacement for local representation.
  • I. INTRODUCTION: ScopeMamba-YOLO combines high-resolution scale reallocation, anisotropic feature reading, off-path selective scanning, and scale-adaptive distributional regression.Its selective-scanning principle keeps contextual modeling separate from the fine-grained convolutional stream.
  • I. INTRODUCTION: The framework includes AMS for resource-aware anisotropic extraction and SA-DFL for reallocating regression support across detection scales.The reported design spans four scalable variants and targets an accuracy–parameter trade-off.

A. Small Object Detection in Aerial Imagery

Aerial small-object detection commonly reallocates computation toward higher-resolution features because tiny targets lose evidence during downsampling. ScopeMamba-YOLO adopts this prior while addressing the resulting loss of contextual reach with efficient long-range modeling.

  • A. Small Object Detection in Aerial Imagery: Existing aerial approaches improve resolution, supervision, or context through slicing, label assignment, feature pyramids, and long-range operators.The paper positions its method within these complementary research directions rather than replacing them wholesale.
  • A. Small Object Detection in Aerial Imagery: Tiny aerial targets often disappear after repeated downsampling, motivating a stride-4 P2 level and removal of the stride-32 P5 stage.The adopted P2/P3/P4 layout uses strides {4, 8, 16} for the 640×640 input.
  • A. Small Object Detection in Aerial Imagery: Many Mamba-based detectors place state-space operators on the main feature path, allowing sparse small-object tokens to share representation capacity with local extraction.This motivates evaluating how selective scanning is integrated with the detail-carrying stream.
  • A. Small Object Detection in Aerial Imagery: Pruning P5 removes roughly 7M parameters but weakens peripheral spatial support for the remaining high-resolution features.The paper treats this as the central resolution–context trade-off to be addressed by later context pathways.

2) Data flow:

The architecture preserves a convolutional high-resolution path while adding anisotropic feature extraction and low-resolution selective-scan context streams. Its neck fuses multiscale features through SS-PAN, and SA-DFL performs scale-adaptive regression.

  • 2) Data flow:: The Context-Cascaded Backbone uses P2/P3/P4 convolutional stages, AMS blocks, relocated SPPF, and two gated context streams that re-enter the main path.The network therefore maintains a high-resolution detection layout without a stride-32 stage.
  • 2) Data flow:: AMS splits channels into local, horizontal-strip, vertical-strip, and identity branches before cross-directional gating, fusion, and residual addition.Adaptive strip branches blend short and long kernels using global-content gates.
  • 2) Data flow:: Four-directional non-causal selective scanning aggregates evidence from the entire feature map at O(HW) cost.Rows and columns are scanned in both directions and their outputs are merged.
  • 2) Data flow:: SS-PAN combines top-down DySample fusion and directional context injection with bottom-up zero-gated selective-scan branches.The neck propagates semantic information toward high-resolution features while retaining selective-scan fusion at P3 and P4.
  • 2) Data flow:: SA-DFL assigns per-scale distributional bin budgets K=[24, 16, 10] for scale-adaptive box regression.The head reallocates regression support across the three resulting feature maps.

2) The placement principle:

Selective scanning is kept off the main detail-carrying path and introduced through reduced-cost, zero-initialized side branches. This placement preserves the convolutional initialization while enabling progressive contextual modulation.

  • 2) The placement principle:: Directly embedding selective scanning in the backbone reduces mAP50 from 0.4949 to 0.4851 at unchanged 47.6 GFLOPs.The matched negative result motivates an alternative integration strategy.
  • 2) The placement principle:: The off-path formulation uses side-branch modulation so the convolutional stream retains the fine-grained representation needed for small-object localization.The scan complements rather than replaces the main-path feature transformation.
  • 2) The placement principle:: Backbone context branches operate on 4×-pooled 32-channel features, reducing token count by 16× and adding 0.44 GFLOPs across two stations.The neck instead narrows the branch width because its scan operates on a full-resolution fusion map.
  • 2) The placement principle:: AMS provides the anisotropic alternative to isotropic 3×3 stacks, using adaptive strip kernels to reduce high-resolution extraction cost.Its four-branch design covers elongated structures while preserving the unchanged C2f shell.
  • 2) The placement principle:: Zero-initialized gates make the augmented network functionally identical to its convolutional skeleton at initialization.This permits contextual pathways to be introduced progressively during optimization.

3) Station 1: Cascaded Global-Context Module (CGCM):

CGCM computes contextual streams from downsampled P3 and P4 features, then injects them through zero-initialized multiplicative gates while preserving the main convolutional pathway.

  • CGCM: The P3 context stream pools by four, projects to 32 channels, applies LayerNorm and selective scanning, then restores the stage width.At 640 × 640 input, the pooled P3 grid is 20 × 20.
  • CGCM: Zero-initialized gating injects aligned context multiplicatively while starting as the identity, preserving the original feature stream at initialization.The gate spans (0, 2) because it uses 1 + tanh(W0g).
  • CGCM: CGCM derives g3 from pooled P3 features and g4 from already-modulated P4 features, accumulating scene context across backbone depth.The cascaded design applies contextual modulation before P4 and again before SPPF.
  • Relation to SS-PAN: SS-PAN complements CGCM with top-down directional context injection toward P2 and bottom-up selective-scan fusion nodes.The top-down pathway uses DySample and directional depthwise strip operators for semantic propagation.

2) Bottom-Up Selective-Scan Fusion:

The bottom-up SS-PAN leaves the C2f main path intact and adds zero-gated selective-scan branches, while SA-DFL assigns scale-specific regression support and capacity.

  • Bottom-up selective-scan fusion: The bottom-up fusion node combines unchanged C2f features with a zero-initialized selective-scan branch: y = C2f(x) + γ ΦSSM(x), γ = 0 at init.The branch compresses, normalizes, scans with NCSSD2D, refines depthwise, and expands back to the output width.
  • Bottom-up selective-scan fusion: Selective-scan branches at P3 and P4 provide global recalibration for concatenated multi-scale features without replacing local fusion.Removing a scan node or tapering scan width reduces accuracy, while tapering provides no parameter advantage.
  • Selective-scan operator: The shared NCSSD2D operator performs non-causal four-directional scanning with linear complexity in token count.It scans rows and columns in both directions and merges the four outputs.
  • Scale-adaptive distributional regression: SA-DFL assigns each detection scale its own bin budget, changing supported regression range and reallocating regression-branch capacity across levels.The design adapts distributional support and capacity rather than changing adjacent-bin spacing.
  • Scale-adaptive distributional regression: 0.008M additional parameters accompany the primary SA-DFL configuration's positive accuracy gain in the cumulative ablation.The head predicts per-scale distributions decoded by expectation.

F. Model Scaling

ScopeMamba-YOLO scales a shared topology through YOLOv8-style depth and width factors, with Nano using a separate SA-DFL bin list to preserve proportional head widths.

  • Scaling design: All four variants share one topology; only compound depth and width factors change, with an additional Nano-specific SA-DFL bin list.Parameters and GFLOPs are reported at 640 × 640 under the 10-class VisDrone configuration.
  • Scaling design: YOLOv8 compound scaling multiplies per-layer channel widths by w and C2f-family repeat counts by d.Widths are rounded to multiples of eight.
  • Nano scaling: Nano uses K=[16, 12, 8] rather than [24, 16, 10], restoring proportional regression widths from [96, 64, 40] to.The tested Nano allocation performs best among the evaluated Nano-scale bin choices.
  • Evaluation setup: Experiments evaluate VisDrone-2019 and AI-TOD using mAP50, mAP50−95, parameter count, and FLOPs, with YOLOv8s as the baseline detector.VisDrone and AI-TOD use official validation splits under the stated training protocol.

B. Comparisons with State-of-the-Art Methods

Across VisDrone-2019 and AI-TOD, ScopeMamba variants show favorable accuracy–parameter trade-offs, with especially strong gains for very-tiny and tiny objects.

  • VisDrone-2019: ScopeMamba variants consistently improve mAP50 over YOLOv8 entries while using substantially fewer parameters across evaluated model scales.The accuracy–parameter relationship is also visualized in Fig. 6.
  • VisDrone-2019: 10.8 pp higher mAP50 and 3.57M parameters make ScopeMamba-S substantially more accurate and smaller than YOLOv8s on VisDrone-2019.ScopeMamba-S reaches 0.508 mAP50 versus YOLOv8s, whose parameter count is 11.10M.
  • VisDrone-2019: 0.526 mAP50 with 6.48M parameters lets ScopeMamba-M exceed reported accuracy for several UAV-specific detectors while using fewer parameters than most.The comparison includes Drone-DETR, SFBF-YOLO-m, and SOD-YOLO-l.
  • AI-TOD: The AI-TOD comparison covers overall accuracy, size-specific AP and AR, model complexity, and computational cost.Representative qualitative comparisons show dense scenes with very small targets.
  • AI-TOD: On AI-TOD, ScopeMamba-S improves AP vt by 4.57 pp and AP t by 4.58 pp over YOLOv8S, while AP s improves by 0.41 pp.The larger gains occur in the very-tiny and tiny regimes.

C. Integration Strategy for Selective Scanning

The study preserves local detail by keeping selective scanning off the main path while using it to restore broader contextual and peripheral support. This design improves spatial reach without attributing gains to auxiliary capacity alone.

  • Main-Path vs. Off-Path Integration: Direct main-path selective scanning reduces mAP50 from 0.4949 to 0.4851 at unchanged 47.6 GFLOPs, motivating off-path integration.The comparison uses an earlier matched P2/−P5 configuration with the default head.
  • Contextual Reach: CGCM gate responses extend across broad scene structures rather than remaining confined to localized object regions.This spatial behavior is consistent with CGCM’s intended role as a scene-level context modulator.
  • Operator Bracket: Selective scanning reaches 0.5078 mAP50, exceeding the cost-matched convolutional branch by 0.63 pp over the three-seed no-branch mean.The convolutional branch changes mAP50 by only +0.04 pp, whereas selective scanning improves it by +0.67 pp.
  • ERF Analysis: The ERF probe averages absolute input-gradient magnitudes from the shared stride-8 P3 center cell across 200 resized VisDrone images.PER measures the fraction of ERF mass outside the central 50% × 50% image region.
  • Contextual Reach: PER rises from 0.008 to 0.090 with the complete off-path context pathway, an approximately 11× increase over the P2/−P5 baseline.ScopeMamba-S recovers approximately 61% of YOLOv8s’s measured peripheral ERF energy while retaining the high-resolution scale allocation.

D. Ablation Study

The cumulative ablation shows that reallocating detection scales toward higher resolution provides the largest accuracy gain, while later modules improve efficiency or add smaller complementary gains. Overall, the final trajectory substantially improves mAP50 while reducing parameters.

  • Cumulative Trajectory: The cumulative ablation reports mAP50, parameter count, and GFLOPs at every integration step, with Fig. 11 mapping performance gains against capacity reductions.Table V provides the stepwise numerical trajectory.
  • Cumulative Trajectory: mAP50 increases by 9.54 pp while parameters decrease approximately 68%, from 11.14M to 3.57M, across the full trajectory.The HRR prior contributes 7.93 pp, while the subsequent redesign contributes another 1.61 pp and reduces parameters by approximately 0.57M.
  • High-Resolution Reallocation: Replacing P5 with P2 raises mAP50 from 0.4124 to 0.4917, a 7.93 pp gain, while reducing parameters from 11.14M to 4.14M.Computation increases from 28.45 to 51.10 GFLOPs because of the higher-resolution feature maps.
  • AMS Block: The AMS Block reduces parameters from 4.14M to 3.13M and GFLOPs from 51.10 to 46.54, while mAP50 changes by −0.07 pp.Anisotropic strip kernels reach 0.5078 mAP50 versus 0.4963 for comparable isotropic square kernels.

3) Effect of SS-PAN (Reach):

SS-PAN and SA-DFL provide complementary scale-aware improvements: selective-scan branches help across evaluated model sizes, while SA-DFL adds accuracy with minimal parameter cost. Additional regression variants did not consistently outperform the adopted design.

  • SS-PAN Contribution: Removing SS-PAN selective-scan branches lowers mAP50 by 0.54 pp, 1.09 pp, and 0.27 pp for the Nano, Small, and Medium variants.The consistent direction indicates positive contributions across the three evaluated scales.
  • SA-DFL Contribution: SA-DFL adds 0.008M parameters and increases mAP50 from 0.5018 to 0.5078, a 0.60 pp gain.Its main resource cost is computational rather than parametric.
  • SA-DFL Contribution: The scale-adaptive reg_max allocation achieves the highest mAP50 among tested Nano configurations by jointly reallocating distributional support and regression capacity.Regression-branch width scales with 4K_i, so bin allocation changes capacity across feature levels.
  • SA-DFL Contribution: Alternative quality-branch and NWD variants did not provide consistent improvements over the adopted SA-DFL configuration in the tested settings.These exploratory controls are supporting negative results, not evidence that SA-DFL is globally optimal.
  • Robustness: The cumulative and multi-seed CGCM comparisons report different effects: +0.45 pp for the single-run trajectory versus +0.67 pp over the three-seed no-CGCM mean.The observed 0.34 pp standard deviation is an empirical reference for the no-CGCM control, not a formal significance estimate for all variants.

F. Discussion on Negative Results

The negative-result controls favor the adopted off-path selective-scanning formulation, SA-DFL regression, and anisotropic feature extraction over tested alternatives. Qualitative comparisons also show ScopeMamba-S detecting more annotated targets than YOLOv8s in challenging VisDrone scenes.

  • Placement and Dimensional Constraints: Direct main-path scanning reduces mAP50 by 0.98 pp, and halving neck scan width reaches 0.4289 versus 0.4326 without the branch while using more parameters.The tapered configuration improves neither accuracy nor parameter efficiency.
  • Operator Substitution: Among tested operators, selective scanning reaches 0.5078 mAP50 at the neck, versus 0.5013 for EMA, while convolution reaches 0.5015 at the backbone.At the backbone, the cost-matched convolution is close to the three-seed no-branch mean of 0.5011, whereas EMA reaches 0.4979.
  • Quality-Branch Stacking and Loss Shaping: The tested quality branches reach 0.4987 and 0.4983, both about 0.9 pp below SA-DFL at 0.5078, while specialized losses reach 0.4980 and 0.4973.NWD changes mAP50 by only +0.04 pp while reducing mAP50−95 by 0.31 pp, indicating a metric trade-off rather than consistent localization improvement.
  • Capacity Reallocation Controls: Reallocating parameters toward depth reduces Small-scale performance by 0.43 pp, while isotropic kernels reduce accuracy by 1.15 pp versus anisotropic AMS kernels.These controls favor context injection and directional selectivity over additional depth or raw kernel extent.
  • Summary: The controls favor off-path selective scanning, scale-adaptive regression, and anisotropic extraction over direct insertion, extra branches, loss shaping, and added depth.The summary identifies the adopted design choices as better supported than the tested alternatives.
  • Qualitative Comparison: ScopeMamba-S detects more annotated pedestrians and vehicles than YOLOv8s in representative dense, occluded, and low-illumination scenes.The highlighted examples include targets missed by the baseline under crowding, weak local appearance cues, and low-light conditions.
Loading 2609.10156v1…