Focus-and-Detect: A Small Object Detection Framework for Aerial Images

1 Introduction

• Aerial images combine small, densely placed objects with substantial scale variation, orientation differences, and class imbalance.VisDrone images may contain up to 902 objects, and car-size variance is reported as almost five times that of MS COCO.Source: §1 · p. 2
• The two-stage framework uses a GMM-supervised detector to locate focal regions, then predicts objects within those regions and merges their predictions.NMS and IBS are applied after merging to address overlapping and truncated boxes.Source: §1 · p. 2, Figure 1 · p. 2
• The proposed GMM method generates scale-normalized object clusters for use as focal regions.The focus network is supervised by cluster coordinates generated using a Gaussian Mixture Model.Source: §1 · p. 2, Figure 1 · p. 2
• IBS suppresses incomplete boxes caused by overlapping focal regions.Source: §1 · p. 2
• 42.06 AP on VisDrone validation and 54.16 AP@70 on UAVDT test; the authors report state-of-the-art small-object performance on VisDrone.The reported comparison is to small-object detection methods in the literature, to the authors’ knowledge.Source: §1 · p. 3

2.1 Object Detection

• Object detectors combine backbones, necks, and heads; small-object performance remains challenging, including for detectors effective on medium or large objects.The review contrasts proposal-based two-stage detectors with dense single-stage detectors and notes that objects under 32 × 32 pixels are considered small in the cited work.Source: §2.1 · p. 3

2.2 Small Object Detection

• Prior small-object approaches address limited image information, scale, or proposal quality through super-resolution, image pyramids, multi-scale features, and region-proposal changes.The review notes that image-pyramid methods are not efficiently scalable.Source: §2.2 · p. 3

2.3 Object Detection in Aerial Images

• Aerial-image methods address scale variation, class imbalance, and dense objects; prior region-search approaches include tiling, clustering, cropping, and density maps.Focus&Detect proposes GMM for region search and IBS for incomplete boxes in overlapping regions; it contrasts its scale normalization with prior approaches.Source: §2.3 · p. 4

3.1 Overview

• The two-stage framework focuses on object-cluster regions before detection, increasing focal-region resolution at lower computational cost than processing high-resolution images directly.IBS and standard NMS are applied after predictions from the regions are merged.Source: §3.1 · p. 4

3.2 Focus Stage

• The focus stage uses a GFL detector with a ResNet-50 backbone and deformable convolutions to predict focal regions.Deformable convolutions dynamically change the receptive field to improve focal-region feature representation.Source: §3.2 · p. 4

3.2.1 Generating Ground-Truth Boxes of Focal Regions Using Gaussian Mixture Model

• GMM clusters GT-box distance vectors to evenly sampled image-grid points, with Nf = log2(Ngt) + 2 regions per image.The number of regions depends on the image’s number of ground-truth boxes.Source: §3.2.1 · pp. 5–6
• EM fits the GMM, and each cluster becomes the minimum enclosing focal-region box with a 20-pixel gap on every side.These generated regions serve as ground-truth boxes for the focus stage.Source: §3.2.1 · p. 6

3.3 Detection Stage

• The detection stage crops and enlarges GMM focal regions for a dedicated detector; truncated ground-truth boxes are retained when at least 30% lies inside a crop.The detector uses GFL, ResNeXt-101, deformable convolutions, and an FPN.Source: §3.3 · pp. 6–7

3.4 Post Processing

• After regional predictions are merged, IBS and NMS are used to improve the final object-box predictions.IBS addresses incomplete boxes; NMS suppresses overlapping detections.Source: §3.4 · p. 7

3.4.1 Incomplete Box Suppression

• Because truncated and full-box duplicates can evade NMS at low IoU, IBS compares boxes across overlapping focal regions and suppresses boxes matched to clipped neighbor-region boxes.IBS clips boxes from overlapping regions to the focal region under consideration, then suppresses a box when clipped-box IoU exceeds the selected threshold.Source: §3.4.1 · p. 7

3.4.2 Non-max Suppression

• NMS suppresses duplicate predictions from overlapping focal regions by retaining the highest-confidence box when box IoU exceeds 0.5.The 0.5 IoU threshold is used after merging regional detections.Source: §3.4.2 · p. 8

4.1 Implementation Details

• Both stages use Generalized Focal Loss with Feature Pyramid Networks; the focus and detection stages use ResNet-50 and ResNeXt-101, respectively.Focal detections are merged using NMS and IBS; training runs for 24 epochs with scheduled learning-rate reductions.Source: §4.1 · p. 8

4.2 Dataset and Evaluation Metric

• VisDrone2021 provides 6,471 training, 548 validation, and 3,190 test images; evaluation reports COCO AP metrics, including size-specific APS, APM, and APL.AP averages precision over ten IoU thresholds from 0.5 to 0.95.Source: §4.2 · p. 9
• UAVDT contains around 41,000 frames and 840,000 boxes; its car, truck, and bus categories are combined as one vehicle class and scored with VOC AP at IoU 0.7.Source: §4.2 · p. 9

4.3 Results

• On VisDrone test data, Focus&Detect improves pedestrian, person, and bicycle scores, while the baseline has higher mAP@50 for tricycle and awning-tricycle.Table 1 compares class-wise AP@50 scores for GFL and Focus&Detect.Source: §4.3 · p. 9, Table 1 · p. 9
• EIP performs worst among the compared region-search methods despite generating many regions; Focus&Detect reports 42.06 AP on VisDrone validation.Table 2 compares region-search methods by image count, precision metrics, and inference time.Source: §4.3 · p. 9, Table 2 · p. 10
• 42.06 AP and 44.64 AP75 are the best reported scores among state-of-the-art methods on VisDrone validation.The comparison reports results from the original papers; Table 3 presents the method comparison.Source: §4.3 · p. 9, Table 3 · p. 10
• Focus&Detect surpasses reported methods on UAVDT, where very small objects are unannotated and ignored during testing.The UAVDT comparison uses the test set; its dataset does not focus solely on small objects.Source: §4.3 · pp. 9–10, Table 4 · p. 10

4.4 Ablation Study

• GMM clustering normalizes box areas and raises their mean, which the authors associate with improved small-object detection; resizing and multi-scale training also normalize areas, unlike resolution increases alone.Figure 6 shows the clustering effect; the authors describe multi-scale training as shifting and normalizing box-area means.Source: Figure 6 · p. 11, §4.4 · p. 12

• Focus&Detect uses GMM to generate focal regions and IBS to mitigate truncated boxes when target regions are merged.Source: §5 · p. 12
• The authors report that IBS improves detection across classes, especially small-object classes, while GMM normalizes object scales across regions and increases overall performance.Source: §5 · p. 12
• The framework achieves state-of-the-art performance on VisDrone validation and UAVDT test among reported small-object detection methods, with the best APS score.Source: §5 · p. 12

• Koyun and Keser are credited with conceptualization, methodology, and original-draft writing.Source: CRediT authorship contribution statement · p. 13
• Keser is additionally credited with data analysis and interpretation.Source: CRediT authorship contribution statement · p. 13
• Akkaya and Töreyin are credited with supervision and review-and-editing.Source: CRediT authorship contribution statement · p. 13