Implementation:NVIDIA DALI EfficientDet Preprocessor

Knowledge Sources	NVIDIA_DALI
Domains	Object_Detection, TensorFlow
Last Updated	2026-02-08 16:00 GMT

Overview

Provides image and bounding box preprocessing functions for EfficientDet detection training, including random horizontal flipping, keypoint manipulation, and multi-scale resizing.

Description

This module implements preprocessing operations for object detection that are applied to images, bounding boxes, instance masks, and keypoints. It distinguishes between operations applied universally (both training and evaluation) and augmentation operations applied only during training.

The core augmentation function is `random_horizontal_flip()`, which randomly mirrors images and their associated annotations with 50% probability. It correctly transforms bounding box coordinates by flipping x-coordinates around the image center, mirrors instance masks along the width axis, and permutes keypoints according to a provided flip permutation (e.g., swapping left_eye and right_eye). All coordinate transformations preserve the normalized [0, 1] coordinate space.

Additional utility functions include `_flip_boxes_left_right()` for deterministic box flipping, `_flip_masks_left_right()` for mask flipping, and `keypoint_flip_horizontal()` for keypoint flipping with permutation. The module also provides `resize_to_range()` for resizing images (and optionally masks) such that the smallest dimension meets a minimum size constraint while the largest dimension does not exceed a maximum, with support for both static and dynamic shape computation. Helper functions `_compute_new_static_size()` and `_compute_new_dynamic_size()` handle the two cases respectively.

The module also includes `box_list_scale()` for scaling box coordinates and various box list utility functions used during the detection preprocessing pipeline. All operations use TensorFlow 1.x compatibility APIs (`tf.compat.v1`).

Usage

Use these functions within the EfficientDet data loading pipeline to apply training augmentations to images and annotations. The functions are called by `DetectionInputProcessor` in the dataloader module during dataset construction.

Code Reference

Source Location

Repository: NVIDIA_DALI
File: docs/examples/use_cases/tensorflow/efficientdet/pipeline/tf/preprocessor.py
Lines: 1-455

Signature

def random_horizontal_flip(
    image,
    boxes=None,
    masks=None,
    keypoints=None,
    keypoint_flip_permutation=None,
    seed=None,
) -> tuple:
    ...

def _flip_boxes_left_right(boxes) -> tf.Tensor:
    ...

def _flip_masks_left_right(masks) -> tf.Tensor:
    ...

def keypoint_flip_horizontal(keypoints, flip_point, flip_permutation, scope=None) -> tf.Tensor:
    ...

def resize_to_range(
    image,
    masks=None,
    min_dimension=None,
    max_dimension=None,
    method=tf.image.ResizeMethod.BILINEAR,
    align_corners=False,
    pad_to_max_dimension=False,
) -> tuple:
    ...

def _compute_new_static_size(image, min_dimension, max_dimension) -> tf.Tensor:
    ...

def _compute_new_dynamic_size(image, min_dimension, max_dimension) -> tf.Tensor:
    ...

Import

from pipeline.tf import preprocessor

# Apply random horizontal flip
image, boxes = preprocessor.random_horizontal_flip(image, boxes=boxes)

I/O Contract

Inputs

Name	Type	Required	Description
image	tf.Tensor	Yes	Rank 3 float32 tensor [height, width, channels]
boxes	tf.Tensor	No	Rank 2 float32 tensor [N, 4] with normalized [ymin, xmin, ymax, xmax] coordinates
masks	tf.Tensor	No	Rank 3 float32 tensor [num_instances, height, width] instance masks
keypoints	tf.Tensor	No	Rank 3 float32 tensor [num_instances, num_keypoints, 2] in y-x normalized coordinates
keypoint_flip_permutation	tf.Tensor	No	Rank 1 int32 tensor specifying keypoint index remapping after flip
seed	int	No	Random seed for reproducibility
min_dimension	int	No	Minimum dimension for resize_to_range
max_dimension	int	No	Maximum dimension for resize_to_range

Outputs

Name	Type	Description
image	tf.Tensor	Preprocessed image tensor (same rank as input)
boxes	tf.Tensor	Transformed bounding boxes (if provided)
masks	tf.Tensor	Transformed instance masks (if provided)
keypoints	tf.Tensor	Transformed keypoints (if provided)

Usage Examples

Apply Training Augmentations

from pipeline.tf import preprocessor

# Random horizontal flip with boxes
image, boxes = preprocessor.random_horizontal_flip(
    image, boxes=boxes, seed=42
)

# Resize to range while preserving aspect ratio
image, masks = preprocessor.resize_to_range(
    image,
    masks=masks,
    min_dimension=640,
    max_dimension=1024,
)

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment