Implementation:NVIDIA DALI Ops Util Normalize Flip

Knowledge Sources	NVIDIA DALI
Domains	Object_Detection, GPU_Computing
Last Updated	2026-02-08 00:00 GMT

Overview

Concrete functions for normalizing images with coordinated horizontal flipping and random crop-resize for object detection, provided by the DALI EfficientDet example utility module.

Description

The ops_util normalize_flip and random_crop_resize functions implement the spatial augmentation stage of the EfficientDet DALI pipeline.

normalize_flip performs two operations in a single step:

1. Generates a random coin flip with probability p (0.5 for training, 0.0 for evaluation).
2. Applies dali.fn.crop_mirror_normalize to subtract ImageNet channel means, divide by standard deviations, and conditionally mirror the image horizontally.
3. Applies dali.fn.bb_flip with the same flip flag to transform bounding box coordinates consistently.

random_crop_resize performs scale-aware cropping:

1. Samples a random scale factor from the given range (e.g., [0.1, 2.0] for training, None for evaluation which uses 1.0).
2. Computes the resize dimensions to fit the scaled image proportionally to the target output size.
3. Resizes the image using dali.fn.resize.
4. Applies dali.fn.random_bbox_crop to jointly crop the image region and transform bounding boxes with the xyXY layout, discarding boxes that fall outside.
5. Extracts the crop region with dali.fn.slice using out-of-bounds padding.

Usage

These functions are called sequentially inside the DALI pipeline definition, after input reading and optional GridMask augmentation.

Code Reference

Source Location

• Repository: NVIDIA DALI
• File: docs/examples/use_cases/tensorflow/efficientdet/pipeline/dali/ops_util.py

Signature

def normalize_flip(images, bboxes, p=0.5):
    ...

def random_crop_resize(
    images, bboxes, classes, widths, heights, output_size, scaling=[0.1, 2.0]
):
    ...

Import

from pipeline.dali import ops_util

# Called as:
images, bboxes = ops_util.normalize_flip(images, bboxes, p=0.5)
images, bboxes, classes = ops_util.random_crop_resize(
    images, bboxes, classes, widths, heights, output_size, scaling
)

I/O Contract

Inputs

Name	Type	Required	Description
images	DALI TensorList	Yes	Decoded RGB images (uint8 or float32).
bboxes	DALI TensorList	Yes	Bounding boxes in normalized ltrb format [N, 4].
p	float	No	Probability of horizontal flip. Defaults to 0.5. Set to 0.0 for evaluation.
classes	DALI TensorList	Yes (crop_resize)	Class labels [N] for filtering during crop.
widths	DALI TensorList	Yes (crop_resize)	Original image widths as float scalars.
heights	DALI TensorList	Yes (crop_resize)	Original image heights as float scalars.
output_size	tuple(int, int)	Yes (crop_resize)	Target (height, width) for the output image.
scaling	list[float] or None	No	Range [min_scale, max_scale] for random scale sampling. None means no scaling (factor = 1.0).

Outputs

Name	Type	Description
images (normalize_flip)	DALI TensorList	Normalized float32 images in NHWC layout, optionally flipped.
bboxes (normalize_flip)	DALI TensorList	Bounding boxes with horizontal flip applied, [N, 4] in ltrb format.
images (crop_resize)	DALI TensorList	Cropped and resized float32 images of shape [output_size[0], output_size[1], 3].
bboxes (crop_resize)	DALI TensorList	Bounding boxes transformed to the crop window, [M, 4] where M <= N.
classes (crop_resize)	DALI TensorList	Filtered class labels [M] corresponding to retained boxes.

Usage Examples

Training Augmentation

# Inside a @pipeline_def function:
images, bboxes, classes, widths, heights = ops_util.input_coco(...)

# Normalize and randomly flip
images, bboxes = ops_util.normalize_flip(images, bboxes, p=0.5)

# Random crop and resize to 512x512
images, bboxes, classes = ops_util.random_crop_resize(
    images, bboxes, classes, widths, heights,
    output_size=(512, 512),
    scaling=[0.1, 2.0],
)

Evaluation (No Augmentation)

# No flip, no random scaling
images, bboxes = ops_util.normalize_flip(images, bboxes, p=0.0)
images, bboxes, classes = ops_util.random_crop_resize(
    images, bboxes, classes, widths, heights,
    output_size=(512, 512),
    scaling=None,
)

Related Pages

Implements Principle

• Principle:NVIDIA_DALI_Spatial_Augmentation_Detection

Requires Environment

• Environment:NVIDIA_DALI_CUDA_GPU_Environment