Batch processing of DNN with varing ROI

a.a.menezes · May 14, 2020, 7:41am

I want to create a classifier with a batch size of 5.
I created a tensorRT file with the option of batch size 5.
My goal is to classify each box detected by DriveNet using this classifier.
In order to perform inferencing I have the following API:

(	float32_t *	d_outputImage ,
	const dwImageCUDA const	inputImages ,
	uint32_t	numImages ,
	const dwRect *	roi ,
	cudaTextureAddressMode	addressMode ,
	dwDataConditionerHandle_t	obj
)

I plan to pass an array for the inputImages argument just as the DriveNet ncameras example.
My question is that is it possible to give an ROI array in case of the roi argument?
Note: The roi.x and roi.y values are also varing.

shayNV · May 14, 2020, 8:28am

Hi @a.a.menezes,
the API dwDataConditioner_prepareData supports only one ROI per call for all the images in the batch. so it is not possible to pass an array of ROIs to the API.
Unfortunately, there are no releases targeted for DRIVE PX2 anymore. it is possible that in future releases for DRIVE AGX this option can be added. Please consider upgrading to DRIVE AGX if possible.