Large-scale foundation models such as the Segment Anything Model 3 (SAM 3) promise a transition toward open-vocabulary, training-free computer vision. However, their capacity to generalize out-of-distribution to the top-down geometric structures of Earth Observation imagery remains largely unquantified. We present a multi-task empirical evaluation of SAM 3 across remote sensing scene classification, object detection, and instance segmentation, without any fine-tuning. For scene classification, we repurpose its decoupled binary presence head as a standalone zero-shot classifier. For detection and segmentation, we compare textual and visual prompts, where a visual prompt reveals one true instance of the target class in the image being segmented. Text alone performs poorly, while adding a visual prompt raises accuracy sharply, and combining both raises or lowers accuracy depending on the task. We also compare SAM 3's text-only performance against several general-domain, open-vocabulary detection and segmentation models under the same evaluation protocol, and find that it leads all of them on both benchmarks. Its accuracy without a visual prompt nonetheless remains limited, motivating parameter-efficient fine-tuning as a direction for future work.
翻译:暂无翻译