Jens-Duttke/Depth-CoreML-ONNX

Depth models rewritten for Core ML

Depth estimation models from other repositories, rewritten so ONNX Runtime's Core ML execution provider takes the whole graph on a Mac - Neural Engine, GPU and CPU - instead of handing parts of it to the CPU. Oku3D uses them on macOS. Core ML runs only standard ONNX operators, in float16, which each file below was adapted to.

File	Original	License	Rewritten
da-v2-small_fp16_coreml.onnx	onnx-community/depth-anything-v2-small, model_fp16.onnx	Apache-2.0	Scalar-index Gathers on rank-5 tensors written as Slice and Squeeze.
da-v2-base_fp16_coreml.onnx	onnx-community/depth-anything-v2-base, model_fp16.onnx	CC-BY-NC-4.0	As Small, plus normalization inputs past float16's range scaled down.
da-v2-large_fp16_coreml.onnx	onnx-community/depth-anything-v2-large, model_fp16.onnx	CC-BY-NC-4.0	As Base.
midas-v21-small_256_coreml.onnx	julienkay/sentis-MiDaS, midas_v21_small_256.onnx	MIT	Constants and shape arithmetic folded.
midas-dpt-swin2-tiny_256_coreml.onnx	julienkay/sentis-MiDaS, dpt_swin2_tiny_256.onnx	MIT	Fixed at 256x256; window partitions at rank 5 instead of 6; ReduceL2 and Expand written in standard operators; normalization inputs past float16's range scaled down.

The weights are the original's; only the graph around them changes. Checked against the original model on ONNX Runtime's CPU provider, the depth maps agree at a correlation of 0.9999 or better. Each file keeps the license of its original; the CC-BY-NC-4.0 models are for non-commercial use only.