Skip to content

Latest commit

 

History

History
149 lines (113 loc) · 6.67 KB

nvOCDR.md

File metadata and controls

149 lines (113 loc) · 6.67 KB

nvOCDR C API Reference

nvOCDR macro

MAX_CHARACTER_LEN

The maximum length of the character to be recognized in nvOCDR library. Default to be 64.

MAX_PLY_CNT

The maximum number of points of the deteced text area polygons. Default to be 8.

MAX_BATCH_SIZE

The maximum number of images can be feed to nvOCDR at once. Default to be 64

nvOCDR enum

nvOCDRStat

The status of execution of nvOCDR library.

Value Description
SUCCESS Success in execution
FAIL Fail in execution

Device

The data location.

Value Description
CPU The data is stored on CPU
GPU The data is stored on GPU

Notes: Only supports data on GPU now.

DataFormat

The input data format.

Value Description
NCHW The data is stored in shape [batch, channel, height, width]
NHWC The data is stored in shape [batch, height, width, channel]

Notes: Only supports NHWC input now.

OCRNetDecode

The decode method of OCRNet.

Value Description
CTC Decode the OCRNet output following CTC rules
Attention Decode the OCRNet with attention module

nvOCDR meta struct

nvOCDRParam

The parameters to initialize nvOCDR library.

Variable Type Description
input_data_format DataFormat The input data format to be processed in the nvOCDR library
ocdnet_trt_engine_path char* The absolute path to the OCDNet TensorRT engine
ocdnet_binarize_threshold float The threshold value to binarize the OCDNet's output mask
ocdnet_polygon_threshold float The threshold value to filter the polygons based on the confidence score. The confidence score of a polygon is the average value of every pixel's probability in the polygon mask
ocdnet_unclip_ratio float The ratio to unclip the contours of text areas
ocdnet_max_candidate int The maximum number of text area to be produced in one single image from OCDNet
ocdnet_infer_input_shape int32_t[3] The TensorRT engine inference shape of OCDNet. In CHW order
upsidedown bool Enable the nvOCDR to recognize the character that is totally upsidedown
ocrnet_trt_engine_path char* The absolute path to the OCRNet TensorRT engine
ocrnet_dict_file char* The absolute path to the OCRNet character list file
ocrnet_infer_input_shape int32_t[3] The TensorRT engine inference shape of OCRNet. In CHW order
ocrnet_decode OCRNetDecode The decode method of OCRNet

nvOCDRInput

The input data structure of nvOCDR. All the data feed into the library should be wrapped in this struct.

Variable Type Description
shape int[4] The shape of input data. In NHWC format
mem_ptr void* The pointer of the input data.
device_type Device GPU

nvOCDROutputBlob

The single text output struct of nvOCDR, which contains the coordinates of the polygon's vertices that covers a text area and the corresponding decoded text.

Variable Type Description
poly_cnt int32_t The valid length of the array stores the coordinates
polys int32_t[MAX_PLY_CNT] The array stores
mem_ptr void* The pointer of the input data.
device_type Device GPU

nvOCDROutputMeta

The wrapped texts output struct for one inference of nvOCDR, which contains the nvOCDROutputBlob

Variable Type Description
batch_size int32_t The batch size of this inference
text_cnt int32_t[MAX_PLY_CNT] The array stores
text_ptr void* The pointer of the input data

nvOCDR function

nvOCDRp nvOCDR_init(nvOCDRParam param)

The function to initialize the nvOCDR library. It will return a handle of nvOCDR library.

Parameter Type Description
param nvOCDRParam The parameters to initialize nvOCDR library

nvOCDRStat nvOCDR_inference(nvOCDRInput input, nvOCDROutputMeta* output, nvOCDRp nvocdr_ptr)

The function to do OCR. It takes batch images data on GPU as input and store the polygons and texts in the output of those images.

Parameter Type Description
input nvOCDRInput The parameters to initialize nvOCDR library
output nvOCDROutputMeta* The pointer to the output struct
nvocdr_ptr nvOCDRp The handle of nvOCDR library

nvOCDRStat nvOCDR_high_resolution_inference(nvOCDRInput input, nvOCDROutputMeta* output, nvOCDRp nvocdr_ptr, float overlap_rate)

The function to do OCR for high resolution input with a crop-based inference scheme. It takes batch images data on GPU as input and store the polygons and texts in the output of those images. The high resolution input will be cropped to patches with shape of OCDNet input. And then the OCR pipeline will be done on the cropped patches.

Parameter Type Description
input nvOCDRInput The parameters to initialize nvOCDR library
output nvOCDROutputMeta* The pointer to the output struct
nvocdr_ptr nvOCDRp The handle of nvOCDR library
overlap_rate float The overlap ratio of cropped patches

void nvOCDR_deinit(nvOCDRp nvocdr_ptr)

The function to destroy the nvOCDR library resources when you don't need the library.

Parameter Type Description
nvocdr_ptr nvOCDRp The handle of nvOCDR library

Performance tuning guide

Besides the hardware the software running on, the performance(images/sec) of nvOCDR library is influenced a lot by the initial parameters. Here are two tips for better performance

Run with optimal batch size and shape.

When create TensorRT engine, there is a option called optShape. The generated engine will have the best performance with this shape. Thus, one should chose the optimal batch size and spatial shape according to the use cases:

  • For OCDNet, the optimal batch size should be the number of input streams. For example, if there is only one single video stream to be processed in DeepStream pipeline, then batch_size=1 is the best choice.

  • For OCRNet, one could chose the average number of text areas in one single image muliplied with the batch size of OCDNet as the batch size of OCRNet.

  • For general performance concern, the batch size and spatial shape should be muliple of 16.

Set OCDNet post-process parameters.

OCDNet postprocessing takes 20% percent of time. We could change the OCDNet postprocessing for better performance in case it won't bother the accuracy a lot.

  • increase the ocdnet_polygon_threshold to filter out more polygons with lower confidence score.
  • decrease the ocdnet_max_candidates to limit the number of OCDNet output to be post-processed.

Disable upsidedown.

If in the use case, there is not much upsidedown characters, you could disable the upsidedown option. This option will bring in extra 30% compute workload.