AGENTS.md

AGENTS.md

Summary

This is the repository reference for offline NeuralDepth evaluation on stereo datasets. Use it when you need accuracy metrics and saved outputs from host-supplied stereo pairs instead of a live visualizer demo.

Use This Example When

• You need NeuralDepth accuracy evaluation on a dataset.
• You want EPE, bad-pixel metrics, and density reported per scene and in aggregate.
• You need host-supplied stereo pairs sent to a connected device.

Do Not Use This Example When

• You need a live camera demo.
• You need standalone packaging.
• You need generic stereo benchmarking unrelated to NeuralDepth.

Quick Facts

• Category: neural-networks/depth-estimation/neural-depth/host_eval
• Shape: eval
• Primary task: evaluate NeuralDepth variants on stereo datasets
• Entrypoint: main.py
• Standalone path: none
• Frontend: none
• Runs on: host mode only; requires a reachable device and a stereo dataset on disk
• Requires: dataset scenes with im0.png, im1.png, and optional disp0.pfm
• Input: host-loaded stereo image pairs
• Output: saved disparity.png files plus printed metrics
• Models: NeuralDepth device-zoo variants mapped in main.py
• Visualizer / UI: none

Read First

• README.md
• main.py
• utils/utils.py
• utils/arguments.py
• utils/middlebury_download.py

Architecture

• The script walks a dataset folder and finds scenes containing left/right stereo images.
• It maps the requested NeuralDepth variant to a device-zoo model and inference size.
• Host code wraps the stereo images into dai.ImgFrame messages and sends them to a NeuralNetwork node through explicit input queues.
• Predictions are postprocessed into disparity visualizations and metric summaries and then written to the output directory.

Constraints

• This is not a Visualizer example; it writes files and prints metrics.
• The CLI uses --device_ip, not the usual --device.
• Dataset layout is opinionated and effectively Middlebury-style.
• Metric computation is only available when ground-truth disparity exists.

Related Examples

• neural-networks/depth-estimation/neural-depth: use this when you need the live NeuralDepth demo
• neural-networks/depth-estimation/foundation-stereo: use this when you want a heavy host stereo comparison baseline
• depth-measurement/stereo-on-host: use this when you need host stereo comparison rather than NeuralDepth evaluation

Validation

• Run: python3 main.py --dataset <DATASET_DIR> --output <OUTPUT_DIR> --device_ip <IP>
• Success looks like: per-scene outputs are written under the output directory and aggregate metrics print at the end
• Common failure meaning: the dataset layout is wrong, the target device is unreachable, or the chosen NeuralDepth variant does not match the intended evaluation setup