jessolmstead/greycard-sky

greycard sky

The sky prior for greycard, a RAW development engine and editor. The editor downloads this on first use of the Sky mask; nothing here needs to be fetched by hand.

One file, eomt-s-coco-640.onnx: EoMT-S (Kerssies et al., "Your ViT is Secretly an Image Segmentation Model", CVPR 2025), trained on COCO panoptic at 640 px, exported to ONNX by greycard from the published weights at tue-mps/coco_panoptic_eomt_small_640_2x (revision 10f5326). The weights are untouched; the export fixes the input at 640 by 640 and traces the transformers port. Its backbone is Meta's DINOv2 ViT-S, Apache-2.0.

greycard reads the sky class out of the 133 COCO panoptic classes, gates the answer so a frame with no sky gets no mask, refines the outline with SAM 2 and takes the edge with a color-line matte. The rest of the label map (water, mountain, vegetation, ground, buildings) is what the landscape masks will read.

Contract

• Input pixel_values: 1 × 3 × 640 × 640, float32. The picture letterboxed to 640 on its long side, padded with black at the bottom or right, ImageNet-normalized (mean 0.485, 0.456, 0.406; std 0.229, 0.224, 0.225).
• Outputs: the class logits, 1 × 200 × 134 (133 classes plus no-object), and the mask logits, 1 × 200 × 160 × 160, one per query.

Reproducing the file

tools/ai/eomt_export.py in the greycard repository, with the pinned versions in tools/ai/requirements.txt, gives these exact bytes: 96,025,182 bytes, sha256 805ed0fb363784811a7f3979335c2c64a7d40f91ae943326f79a56ff477f372d.

License

MIT, the upstream license, with its notice in LICENSE (copyright 2025 Mobile Perception Systems Lab at TU/e). The export is a format conversion of the published weights and carries no other terms. greycard itself is GPL-3.0-or-later: https://github.com/jessolmstead/greycard.