Tiny Object Detection in Aerial Imagery

Research collaboration · SmartAILab & Vision Lab, Soongsil University (Prof. Seongsin Kim & Prof. Youngjun Han) · 2024

Very small objects (~10–13 px; vehicles, ships, aircraft) in high-resolution aerial imagery often lose discriminative detail as detector feature maps are downsampled. I helped design CAHF [1], a cross-attention module that recovers small-object cues by fusing two complementary feature streams — object detection and segmentation — and evaluated it on the AI-TOD aerial tiny-object benchmark [2].

Cross-attention fusion of heterogeneous feature maps

CAHF architecture: spatial attention and channel cross-attention CAHF fuses object-detection and segmentation features via (a) spatial attention and (b) channel cross-attention.

  • Fuse object-detection and segmentation streams to recover tiny-object cues that are weakened during detector downsampling [1].
  • Spatial attention: reduces and concatenates detection/segmentation feature maps to generate a spatial gate for suppressing background noise.
  • Channel attention: computes cross-task channel correlations and converts them into per-channel weights that emphasize the most informative channels.

Feature-level evidence for tiny-object focus

Feature maps with vs. without CAHF Feature maps before and after CAHF — CAHF enhances foreground responses and suppresses background clutter.

  • Feature-map visualization shows clearer separation between small objects and background after applying CAHF [1].
  • CAHF concentrates responses on small, densely packed objects that a plain detector blurs into the background.

Benchmarking on AI-TOD

Training loss and mAP vs. Faster R-CNN
Training loss and mAP@0.5:0.95 vs. Faster R-CNN — CAHF shows lower loss and consistently higher mAP during training.
Model mAP.5:.95 ↑ mAP.5 ↑
Faster R-CNN 11.1 26.3
Cascade R-CNN 13.8 30.8
M-CenterNet 14.5 40.7
DetectoRS 14.8 32.8
YOLOv8l 14.9 32.5
CAHF (ours) 19.9 41.4
CAHF (ours, 20 ep) 23.1 43.8
On AI-TOD (ResNet-50 backbone) — CAHF outperforms standard and recent detectors.
  • On AI-TOD (aerial tiny-object benchmark [2]), CAHF reaches 23.1 mAP@0.5:0.95 / 43.8 mAP@0.5, improving over the 20-epoch Faster R-CNN baseline and outperforming standard detectors.
  • Training curves show faster, more stable convergence and consistently higher mAP than the Faster R-CNN baseline.
  1. Seungchan Kwon, Doeon Kim, Gyuil Lim, Youngjun Han, Seongsin Kim. “Tiny Object Detection Method Based on Cross Attention of Heterogeneous Feature Maps.” KIIS Autumn Conference, 2024.
  2. J. Wang, et al. “Tiny Object Detection in Aerial Images.” Int. Conf. on Pattern Recognition (ICPR), 2021.
  3. S. Kwon, G. Lim, Y. Han. “SPAR-Det: Segmentation-guided and Prior-Aided Routing for Small Object Detection.” IEEE/CVF Winter Conf. on Applications of Computer Vision (WACV), 2026.