研究目的
To address the wide-scale variation problem of multiclass object detection in optical remote sensing images by proposing an effective object detection framework based on YOLOv2, improving performance for both small and large objects without adding extra parameters.
研究成果
The proposed framework effectively addresses the scale variation problem in multiclass object detection for optical remote sensing images by using a feature introducing strategy with oriented response dilated convolution. It achieves a 4.4% improvement in mAP over YOLOv2 without extra parameters, demonstrating robustness and efficiency for multiscale objects in complex scenes, though efficiency is comparable to YOLOv2.
研究不足
The proposed model has no advantage in efficiency compared to YOLOv2 due to increased feature map size in the last layer, which affects parallel acceleration in steps like region proposal generation and non-maximum suppression. Additionally, the model may not perform well on extremely large objects in some cases, as noted in visual evaluations.
1：Experimental Design and Method Selection:
The study modifies YOLOv2 by designing a network that concatenates feature maps from different depths and uses an oriented response dilated convolution strategy for feature introduction to handle scale variations.
2：Sample Selection and Data Sources:
The DOTA dataset, containing 2806 aerial images with 15 object categories, is used for training and testing. Images are cropped into 1024x1024 patches for processing.
3：List of Experimental Equipment and Materials:
A GTX Titan Xp GPU with 12 GB memory is used for acceleration. Software includes PyTorch
4：0 and DARKNET for YOLOvExperimental Procedures and Operational Workflow:
Networks are pretrained on VOC datasets, trained for 60 epochs with specific learning rates, and evaluated using mAP on the DOTA validation and test sets. Testing involves cropping images, detecting objects, and combining results.
5：Data Analysis Methods:
Performance is measured using mean average precision (mAP), with comparisons to baseline methods like YOLOv2, YOLOv3, and a method from [7].
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
