coreml-projects/DepthPro-coreml-normalized-inverse-depth-pruned-10-quantized-linear

This repo contains DepthProNormalizedInverseDepthPruned10QuantizedLinear.mlpackage(745 MB).

Normaized Inverse Depth means that the model will output values between $[0,1]$ where 1 is the closest pixel to the camera, and 0 is the furthest pixel from the camera.

This model was first pruned to 10% sparsity, then the weights were linearly quantized.

DepthPro CoreML Models

DepthPro is a monocular depth estimation model. This means that it is trained to predict depth on a single image.

DepthPro paper

DepthPro original repo

Model Inputs and Outputs

Inputs

• image: $1536 \times 1536$ 3 color image ($[1 \times 3 \times 1536 \times 1536]$ ImageType).

Outputs

• normalizedInverseDepth 1536x1536 monochrome image ($[1 \times 1 \times 1536 \times 1536]$ ImageType).

Download

Install huggingface-cli

brew install huggingface-cli

To download:

huggingface-cli download \
  --local-dir models --local-dir-use-symlinks False \
  coreml-projects/DepthPro-coreml-normalized-inverse-depth-pruned-10-quantized-linear \
  --include "DepthProNormalizedInverseDepthPruned10QuantizedLinear.mlpackage/*""

To download everything, skip the --include argument.

Conversion Tutorial

The huggingface/coreml-examples repository contains sample conversion code for DepthProNormalizedInverseDepthPruned10QuantizedLinear.mlpackage and other models.

Swift Integration

The huggingface/coreml-examples repository contains sample Swift code for DepthProNormalizedInverseDepthPruned10QuantizedLinear.mlpackage and other models. See the instructions there to build the demo app, which shows how to use the model in your own Swift apps.