DepthART (AX650 Relative Depth Estimation)

DepthART is a family of relative depth estimation models based on selective state spaces (Selective Scan). This repository provides the ONNX model of the small / 448 variant, an axmodel compiled for AX650, and tools for host ONNX inference, board axmodel inference, and model rewriting.

The model outputs affine-invariant relative depth: larger values mean closer, smaller values mean farther. It does not directly produce metric depth; align it with ground truth depth using a global scale/offset before use.

Single-image result

Left: input image; right: relative depth output by the model (INFERNO colormap, stretched to the 2%–98% percentile range).

Repository contents

Path	Description
models/relative_s_448_default.onnx	Original FP32 ONNX model, containing the custom op com.depthart::SelectiveScan
models/relative_s_448_default_ax650.axmodel	Compiled AX650 model
src/infer/depthart_infer_onnx.py	Host ONNX inference (rewrites SelectiveScan automatically)
src/infer/depthart_infer_axmodel.py	AX650 board axmodel inference (standalone script)
src/tool/depthart_export_axonnx.py	Rewrites SelectiveScan into a pure ONNX subgraph that pulsar2 can compile
src/tool/compare_onnx_axmodel.py, compare_depth_outputs.py	Numerical comparison of outputs from different backends
src/tool/evaluate_nyu654_depth.py	Accuracy evaluation on the NYU Depth V2 eigen test split
host.jpg, host_depth.jpg	Single-image example input and side-by-side visualization result

Model interface

ONNX

Item	Value
Input	image [1, 3, 448, 448] float32, NCHW, RGB
Output	depth [1, 1, 448, 448] float32, relative depth (larger = closer)

AX650 board inference

The script src/infer/depthart_infer_axmodel.py does not depend on other scripts in the repository. For deployment, just place the following files in the same directory on the board:

relative_s_448_default_ax650.axmodel
depthart_infer_axmodel.py
images or image directory to test

Board runtime requires axengine.

Single-image validation

python3 depthart_infer_axmodel.py \
    --axmodel relative_s_448_default_ax650.axmodel \
    --image ./host.jpg \
    --output ./results/ax_single

Outputs results/ax_single/host_depth.jpg (side-by-side visualization) and host_depth.npy; the console prints depth statistics, NPU inference time, CPU processing time, and per-stage timings. A single image is the first inference and includes runtime first-execution overhead; use dataset mode for performance evaluation.

Dataset inference

By default, fast mode only measures read / preprocess / npu and skips depth restoration, visualization, and file writing, for performance testing:

python3 depthart_infer_axmodel.py \
    --axmodel relative_s_448_default_ax650.axmodel \
    --dataset ./images \
    --output ./results/board

Full output (depth restoration, visualization, summary NPZ), for accuracy evaluation:

python3 depthart_infer_axmodel.py \
    --axmodel relative_s_448_default_ax650.axmodel \
    --dataset ./images \
    --output ./results/board_full \
    --full-output \
    --dump-npz ./results/board_full/raw_outputs.npz

Image-by-image comparison against the host ONNX baseline summary NPZ:

python3 depthart_infer_axmodel.py \
    --axmodel relative_s_448_default_ax650.axmodel \
    --dataset ./images \
    --output ./results/board_check \
    --ref-npz ./results/onnx_full/raw_outputs.npz