Table 1 Dataset description.
Particulars | VEDAI dataset | ISPRS potsdam dataset |
|---|---|---|
Purpose | Designed specifically for small object detection (vehicles) in aerial imagery | Urban semantic segmentation and object classification, including vehicles |
Data size | 3687 cropped vehicle samples from aerial images | 2244 annotated vehicle instances extracted from 38 large tiles |
Spatial resolution | 12.5Â cm/pixel | 5Â cm/pixel |
Image type | RGB + IR (infrared) | True orthophoto with 4-band imagery (RGB + NIR) |
Scene type | Rural, urban, and industrial zones | Dense urban scenes with diverse infrastructure |
Object scale | Small object instances (vehicles ~ 20–50 pixels) | Small to medium vehicle sizes with occlusion and overlap |
Annotations | Bounding boxes per object (class-labeled) | Pixel-level semantic masks with object IDs and classes |
Vehicle classes | 9: Car, Truck, Van, Pickup Car, Boat, Camping Car, Other, Plane, Tractor | 4: Car, Truck, Van, Pickup Car |
Relevance | Ideal for training models for fine-grained vehicle type classification under varying environments | Represents complex urban environments with dense traffic and shadows—ideal for real-world deployment evaluation |
Preprocessing applied | Cropping to fixed-size patches, normalization, data augmentation (rotation, flipping, brightness), class balancing | Patch-based tiling from large orthophotos, normalization, data augmentation, class balancing |
