Examples

The examples are provided in the packages/dali_pipeline_framework/examples/ directory.

Pipeline Setup Examples

Example	Description	Entry Point Script
2D Object Detection Pipeline	Shows how to set up a pipeline for 2D object detection. This example also demonstrates the flexibility of the pipeline framework by allowing the switch between single-camera and multi-camera data (adding a hierarchy level to the input data format), where no changes are needed in the pipeline implementation.	packages/dali_pipeline_framework/examples/example_pipeline_obj_det_2d.py
StreamPETR Pipeline	Demonstrates a pipeline for StreamPETR, a 3D object detection model which uses multi-camera data. Note that this pipeline can be directly used as a drop-in replacement for a PyTorch DataLoader in StreamPETR training.	packages/dali_pipeline_framework/examples/example_pipeline_stream_petr.py
Flexible Step Implementation	Demonstrates how to implement a flexible step which can be used to process data in a flexible way. The ImageDecoder class is implemented here, starting from a simple step with hard-coded input data format, and then modified to achieve the flexibility of automatically finding and processing all relevant images in the input data.	(No script)

Data Loaders for NuScenes

Both of the pipeline examples above rely use the NuScenes dataset. To enable the use of the dataset in the corresponding pipelines, related data loaders are provided. These data loaders share a large part of the functionality (both load the same data from NuScenes, but process and output it in a different format). Therefore, they share a large part of the implementation by using helper classes to read the data and perform the necessary conversions. See Data Loading for NuScenes.