No training, no change labels: SpotDiff compares pixel-level features to reach 82.46 F1 on remote sensing change detection
Synopsis
Remote sensing change detection usually depends on large collections of temporally aligned, pixel-level annotations, which limits scaling across regions and platforms. SpotDiff composes DINOv3 semantic features, AnyUp pixel-resolution feature reconstruction and SAM 3 text-guided region proposals into a training-free pipeline, then compares the cosine similarity of the two dates' reconstructed features inside the proposed regions. On LEVIR-CD the ViT-7B version reaches 82.46 F1, 2.06 points above the strongest prior training-free method; on SECOND it raises building-change IoU by 10.3.
Interpretation
SpotDiff removes the need for any change labels or fine-tuning: given two co-registered images of the same area from different dates and an optional text prompt naming a category, it outputs a binary change mask for that category. Earlier training-free methods often localize fine-grained changes poorly because their features are coarse or inconsistent across time, while supervised methods need large collections of temporally aligned pixel-level annotations that are rarely available across regions, imaging platforms and application domains. On LEVIR-CD, SpotDiff with ViT-7B reaches 82.46 F1 and 70.15 IoU, 2.06 F1 and 2.95 IoU above OmniOVCD and 14.46 F1 above AdaptOVCD; the ViT-L version reaches 80.10 F1, on par with OmniOVCD.
The pipeline separates where change might be from whether it actually happened: the segmentation model only proposes candidate objects, and the decision is made by the mean similarity of the two dates' semantic features inside each object. DINOv3 emits patch-level features, one vector per 16×16 pixels, and comparing them directly gives blurry similarity maps; AnyUp is a universal feature upsampler that reconstructs them at pixel resolution guided by the image, whereas prior feature-comparison pipelines mostly used bilinear interpolation. On LEVIR-CD, replacing AnyUp with bilinear interpolation drops F1 from 82.46 to 68.69 and with nearest neighbor to 66.88, so the universal upsampler alone adds 13.8 F1; SWIMS multi-scale sliding-window inference lifts ViT-7B F1 from 61.09 to 82.46.
Perspective
The demonstrated scope is co-registered RGB optical imagery: quantitative evaluation on four public benchmarks plus qualitative transfer to PlanetScope 3 m and Sentinel-2 10 m imagery; at 10 m, isolated buildings span only a few pixels and are detected as small blobs, which marks the practical lower limit of the approach.
The similarity threshold is fixed, and the authors suggest adaptive thresholding from image statistics could improve robustness; accuracy depends on the text prompt, with F1 varying by up to 5.9 points across four synonymous prompts; region priors help most for object-like categories and less for amorphous ones such as impervious surface; multispectral and SAR inputs are left to future work.