agrippa-stereocam

depth-capture

Capture a single stereo frame, compute disparity using StereoSGBM, and write an RGBA PNG where RGB channels carry the camera colour and the alpha channel encodes normalised depth.

The output is center-cropped to a square and sized to a multiple of 32 pixels, ready for direct consumption by ML training pipelines.

Examples

ag-cam-tools depth-capture -a 192.168.0.201 --calibration-local calibration/calibration_20260225_143015_a1b2c3d4 --max-depth 500
ag-cam-tools depth-capture -a 192.168.0.201 --calibration-slot 0 --max-depth 300 --size 640
ag-cam-tools depth-capture -a 192.168.0.201 -A --calibration-slot 0 --max-depth 500 -o ./depth_frames
ag-cam-tools depth-capture -a 192.168.0.201 --calibration-local ./calib --max-depth 200 -b 2

Options

Option	Description
-s, --serial	Match camera by serial number
-a, --address	Connect by camera IP address
-i, --interface	Force NIC selection
-o, --output	Output directory, defaulting to the current working directory
-x, --exposure	Exposure time in microseconds
-g, --gain	Sensor gain in dB
-A, --auto-expose	Auto-expose and then lock
-b, --binning	Sensor binning factor: 1 or 2
-p, --packet-size	GigE packet size in bytes
--calibration-local	Calibration session directory on disk (required)
--calibration-slot	On-camera calibration slot: 0, 1, or 2 (required)
--max-depth	Maximum depth for alpha normalisation, in centimetres (required)
--size	Output side length in pixels; must be a positive multiple of 32 (default: auto)
-v, --verbose	Print diagnostic register readback

Output format

The output is a single RGBA PNG file:

• R, G, B — camera colour from the rectified left eye. For grayscale input (e.g. --binning 2), the grayscale value is replicated across all three channels.
• A (alpha) — depth normalised to [0, 255]:
  • alpha = clamp(depth_cm * 255 / max_depth_cm, 0, 255)
  • Invalid or occluded pixels (where disparity yields no depth) get alpha = 0.

Square crop and sizing

The raw rectified image (typically 1440 x 1080) is center-cropped to the largest inscribed square, then resized to a multiple of 32 pixels.

• Default: the output side is floor(square_side / 32) * 32. For a 1080-tall image, this produces a 1056 x 1056 output.
• --size N: overrides the automatic sizing. N must be a positive multiple of 32. The image is resized using nearest-neighbour interpolation.

Calibration requirement

Calibration is mandatory. Supply either --calibration-local or --calibration-slot (mutually exclusive). The calibration must include:

• Stereo rectification remap tables (for disparity computation).
• focal_length_px and baseline_cm in the calibration metadata (for disparity-to-depth conversion).

Sensor modes

depth-capture requires a stereo Lucid camera. Run against a monocular camera (e.g. Triton TRT016S), it exits immediately with non-zero status and:

error: depth-capture requires a stereo Lucid camera; detected monocular sensor on TRT016S-C (243000112)

Notes

• The disparity backend is SGBM (classical). SGBM parameters are seeded from calibration metadata (min_disparity, num_disparities).
• -A is mutually exclusive with explicit -x and -g.
• When the camera reports that post-binning data is no longer Bayer, the tool treats the input as grayscale and replicates it into RGB.

This site is open source. Improve this page.