Qualcomm AI Engine Direct - Add depth_anything_v2_small to oss_scripts

[zhaoxul-qti]

Summary

• Enable depth_anything_v2_small model (https://github.com/DepthAnything/Depth-Anything-V2) on HTP and integrates the script into oss_scripts/.
• Provide --dump_example_output flag to dump the example image and export depth‑estimation images from both source model and QNN outputs.
• Add unit tests to cover the added model.

Notes

• SQNR is used for matrix evaluation as it provides a more intuitive comparison for our use case compared to AbsRel or RMSE referenced in the paper.
• The model accepts input images of arbitrary resolutions. For evaluation, we use ImageNet samples, rather than the dataset cited in the original publication.

Constraints

• The model is currently unsupported on GPU backend due to the unsupported cubic mode of Resize op.
• GPU support will be enabled once constant folding is introduced, allowing the cubic Resize op to be eliminated.

Test plan

• Test with random images from ImageNet:
  python examples/qualcomm/oss_scripts/depthanything_v2_small.py -a $ARTIFACT -d $IMAGENET_FOLDER_PATH -b build-android/ -H $HOST_NAME -s $DEVICE_ID -m $SOC_ID --seed 1126

• Test with the example image and export the post-processed source model output and QNN output into depth-estimation images:
  python examples/qualcomm/oss_scripts/depthanything_v2_small.py -a $ARTIFACT -d $IMAGENET_FOLDER_PATH -b build-android/ -H $HOST_NAME -s $DEVICE_ID -m $SOC_ID --dump_example_output

Results

Example Image:

Source Model Output (with depth-estimation post-processing):

QNN-HTP Output (with depth-estimation post-processing):