ETD

• semantic segmentation

High-resolution imagery acquired by unmanned aerial vehicles presents a challenging semantic segmentation problem because large continuous regions coexist with small objects, strong perspective effects, and substantial within-image scale variation. Reducing a complete UAV frame to a fixed network input preserves global scene context but can remove spatial detail, whereas local high-resolution processing can preserve more local detail while changing the amount of context available to the model.
This thesis investigates how training resolution, inference resolution, and spatial context interact in semantic segmentation of the UAVid dataset. The experimental study first separates training and inference resolution through a controlled grid in which YOLO26M-SEM models are trained at input sizes of 768, 896, and 1024 pixels and evaluated at 768, 896, 1024, and 1152 pixels. The strongest configuration in this grid is obtained by training at 768 and evaluating at 1152, reaching 0.658414 mIoU. Increasing inference resolution improves all three trained models within the tested range, while increasing training resolution does not produce the same monotonic behaviour.
A second stage evaluates patch-based training together with sliding-window reconstruction of full-resolution UAV images. A model trained on 512 × 512 patches improves the previous global baseline from 0.670259 to 0.694626 mIoU, and horizontal-flip test-time augmentation increases the result to 0.701403. In contrast, applying sliding-window inference to the globally trained model reduces performance to 0.488332 mIoU, indicating that local inference depends strongly on compatibility between the training and inference contexts.
The final stage performs resolution-specific fine-tuning at 768 and 1024 from the same fixed-512 semantic initializer. Although the 768 branch achieves the higher pooled validation score, the completed 1024 epoch-7 configuration with horizontal-flip test-time augmentation and K5 local refinement obtains the highest official-test mIoU among the completed configurations, reaching 0.693964 with a pixel accuracy of 0.865798.
Overall, the results show that resolution should not be treated as a single fixed preprocessing parameter. Its effect depends on where it is introduced in the training and inference pipeline and on the spatial context seen by the model.