YAML Pipeline Operators

Reference for all pre- and post-processing operators that can be used in the pipeline: section of a network YAML file.

These operators handle the CPU-side work: loading inputs, transforming images before the model, and decoding outputs after it. They are distinct from the GStreamer operators, which run in the live-video pipeline.

Input operators​

input​

Loads the input image.

Setting color_format here can eliminate a separate convert-color step later — the color conversion is absorbed into the load.

input-from-roi​

Loads a region of interest from the input image. Used in cascaded pipelines where the first model outputs a bounding box.

Preprocess operators​

Geometric transformations​

centercrop​

Crops the input image from the center. Equivalent to torchvision.transforms.CenterCrop().

letterbox​

Resizes the image while preserving aspect ratio, padding the remainder with a constant value to reach the target dimensions. Used by YOLO models.

resize​

Resizes the input image to specified dimensions.

Pixel value transformations​

contrast-normalize​

Stretches contrast by linearly mapping pixel values to the full available range.

Formula: (input − min) / (max − min)

Use when images have poor contrast and no fixed input range is known.

linear-scaling​

Applies a linear scale-and-shift transformation.

Formula: output = input × scale + shift

Common uses:

normalize​

Standardizes pixel values using per-channel mean and standard deviation.

Formula: output[c] = (input[c] − mean[c]) / std[c]

ImageNet standard values: mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225]

Color space transformations​

convert-color​

Converts the color space of the input image. Follows OpenCV cvtColor conventions.

Supported conversions: RGB2GRAY, GRAY2RGB, RGB2BGR, BGR2RGB, BGR2GRAY, GRAY2BGR.

Tensor conversion​

torch-totensor​

Converts the image to a PyTorch tensor, permuting dimensions to NCHW format.

Set scale: false when you have already scaled values via linear-scaling or normalize — otherwise you get double-scaling.

Postprocess operators​

Mapping training transforms to YAML operators​

When deploying a model, the YAML preprocessing must exactly replicate the transforms used during training. Here are the most common patterns:

Pattern 1 — PyTorch ToTensor()​

transforms.ToTensor()
# or manually: img = img / 255.0; img = np.transpose(img, (2,0,1))

- torch-totensor:
    scale: true   # default, can be omitted

Pattern 2 — Scale to [−1, 1]​

img = img / 127.5 - 1
img = np.transpose(img, (2, 0, 1))

- linear-scaling:
    scale: 0.00784   # 1/127.5
    shift: -1
- torch-totensor:
    scale: false     # values already scaled — do NOT divide by 255 again

Pattern 3 — ImageNet normalization (ToTensor + Normalize)​

transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

- torch-totensor:
    scale: true
- normalize:
    mean: [0.485, 0.456, 0.406]
    std: [0.229, 0.224, 0.225]

Pattern 4 — Offset scaling (−127.5 then × 1/128)​

img -= 127.5
img *= 0.0078125   # 1/128
img = np.transpose(img, (2, 0, 1))

- normalize:
    mean: 127.5
    std: 128
- torch-totensor:
    scale: false   # values already in final range

Key rules​

• torch-totensor always permutes to NCHW — place it last in the preprocess chain.
• torch-totensor with scale: true divides by 255. If you apply linear-scaling or normalize first in integer space, set scale: false.
• normalize expects values in [0, 1] when used after torch-totensor with scale: true. When used before torch-totensor, ensure scale: false.
• Use convert-color when the model expects a different channel order than your input source (e.g. camera outputs BGR, model expects RGB).