## Use Case and High-Level Description¶

ocrnet-hrnet-w48-paddle is a semantic segmentation model, pre-trained on on Cityscapes dataset for 19 object classes, listed in <omz_dir>/data/dataset_classes/cityscapes_19cl_bkgr.txt file. See Cityscapes classes definition for more details. The model was built on HRNet backbone and address the semantic segmentation problem characterizing a pixel by exploiting the representation of the corresponding object class using Object-Contextual Representations. This model is used for pixel-level prediction tasks. For details see repository, paper.

## Specification¶

Metric

Value

Type

Semantic segmentation

GFlops

324.66

MParams

70.47

Source framework

## Accuracy¶

Metric

Value

mean_iou

82.15%

Accuracy metrics were obtained with fixed input resolution 2048x1024 on CityScapes dataset.

## Input¶

### Original model¶

Image, name: x, shape: 1, 3, 1024, 2048, format: B, C, H, W, where:

• B - batch size

• C - number of channels

• H - image height

• W - image width

Expected color order: RGB. Mean values: [127.5, 127.5, 127.5], scale values: [127.5, 127.5, 127.5]

### Converted Model¶

Image, name: x, shape: 1, 3, 1024, 2048, format: B, C, H, W, where:

• B - batch size

• C - number of channels

• H - image height

• W - image width

Expected color order: BGR.

## Output¶

### Original Model¶

Integer values in a range [0, 18], which represent an index of a predicted class for each image pixel. Name: argmax_0.tmp_0, shape: 1, 1024, 2048 in B, H, W format, where:

• B - batch size

• H - image height

• W - image width

### Converted Model¶

Integer values in a range [0, 18], which represent an index of a predicted class for each image pixel. Name: argmax_0.tmp_0, shape: 1, 1024, 2048 in B, H, W format, where:

• B - batch size

• H - image height

• W - image width

You can download models and if necessary convert them into OpenVINO™ IR format using the Model Downloader and other automation tools as shown in the examples below.

omz_downloader --name <model_name>
omz_converter --name <model_name>