DCENet: A Tiny Object Detection Network for Aerial Images Based on Deformable Cross-Attention and Enhanced Classifier

This is the codebase of my journal paper DCENet: A Tiny Object Detection Network for Aerial Images Based on Deformable Cross-Attention and Enhanced Classifier

Model Architecture

Detection stage Weights

Name	Dataset	Input Size	Epochs	mAP	Download(.pth)
DCENet	VisDrone-DET	1920*1920	80	36.0	BaiduYun
DCENet	UAVDT	1024*1024	36	27.5	BaiduYun

Enhanced Classifier Weights

The VisDrone and UAVDT datasets are cropped into numerous image patches according to their ground truth. Subsequently, CSRM is employed to upscale the resolution, thereby forming classification datasets. ResNet-34 is then trained separately on these two classification datasets. The two datasets and model weights are provided below.

Name	Classifier Dataset	Input Size	Epochs	Download(.pth)
ResNet-34	VisDrone-DET	224*224	25	BaiduYun
ResNet-34	UAVDT	224*224	25	BaiduYun