Depth-Anything is a recent advancement in monocular depth estimation which leverages large unlabeled datasets combined with semi-supervised training and DINOv2 as a semantic-encoder.
The authors provide weights of three versions of the model, available on HuggingFace as PyTorch Modules. The repository fabio-sim/Depth-Anything-ONNX has scripts which convert the model weights to ONNX models. Using these ONNX models as the base, I've fused pre/postprocessing operations in it to create a single ONNX model which takes a RGB image as input and outputs a depth-map.
This model is used for inference in an Android app, using onnxruntime's Android libraries, and further applies the Inferno colormap to the depth-map output. Here's the Android app: shubham0204/Depth-Anything-Android
The Android app showcases many abilities of the ONNX and onnxruntime. Inclusion of this sample in onnxruntime-inference-examples would be of great help to app developers.
Depth-Anything is a recent advancement in monocular depth estimation which leverages large unlabeled datasets combined with semi-supervised training and DINOv2 as a semantic-encoder.
The authors provide weights of three versions of the model, available on HuggingFace as PyTorch
Module
s. The repository fabio-sim/Depth-Anything-ONNX has scripts which convert the model weights to ONNX models. Using these ONNX models as the base, I've fused pre/postprocessing operations in it to create a single ONNX model which takes a RGB image as input and outputs a depth-map.This model is used for inference in an Android app, using
onnxruntime
's Android libraries, and further applies theInferno
colormap to the depth-map output. Here's the Android app: shubham0204/Depth-Anything-AndroidThe Android app showcases many abilities of the ONNX and
onnxruntime
. Inclusion of this sample inonnxruntime-inference-examples
would be of great help to app developers.