We build the dim and small vehicle dataset (DSVD) to evaluate the small object detection performance in remote sensing scenes.
For the picked images up to 3225, we labeled the vehicles thoroughly using the rectangular bounding box. 2257 images are randomly chosen from the whole images for the network training, and the other 968 images are the test data for network perfor-mance evaluation. The constructed small object dataset has the following detection difficulties:
• The vehicles with relatively monotonic appearance are small in size, dim in fea-tures, and low in contrast.
• The image covers a wide area and various complex scenes, such as parking lots, roads, neighborhoods, etc.
The dataset is avaliable on the link https://pan.baidu.com/s/1HKBFrgTNDsOInnK0w6qyOg. The extration password is tcjr. And the decompression password is dimandsmallvehicle. If you want use our dataset, please follows these rules:
• All images and their associated annotations in DSVD can be used for academic purposes only, but any commercial use is prohibited.
• If you need to use our dataset to make research, please cite our essay. (The link will be updated when the essay is accepted)
If you have any the problem or feedback in using DSVD, please contact
Tianjun Shi at shitianjun@stu.hit.edu.cn.