# f3net¶

## Use Case and High-Level Description¶

F3Net: Fusion, Feedback and Focus for Salient Object Detection. For details see the repository, paper

## Specification¶

Metric

Value

Type

Salient object detection

GFLOPs

31.2883

MParams

25.2791

Source framework

PyTorch*

## Accuracy¶

Metric

Value

F-measure

84.21%

The F-measure estimated on Pascal-S dataset and defined as the weighted harmonic mean of precision and recall.

F-measure = (1 + β^2) * (Precision * Recall) / (β^2 * (Precision + Recall))

Empirically, β^2 is set to 0.3 to put more emphasis on precision.

Precision and Recall are calculated based on the binarized salient object mask and ground-truth:

Precision = TP / TP + FP, Recall = TP / TP + FN,

where TP, TN, FP, FN denote true-positive, true-negative, false-positive, and false-negative respectively. More details regarding evaluation procedure can be found in this paper

## Input¶

### Original model¶

Image, name - input.1, shape - 1, 3, 352, 352, 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 - [124.55, 118.90, 102.94] Scale values - [56.77, 55.97, 57.50]

### Converted model¶

Image, name - input.1, shape - 1, 3, 352, 352, 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¶

Saliency map, name saliency_map, shape 1, 1, 352, 352, format B, C, H, W, where:

• B - batch size

• C - number of channels

• H - image height

• W - image width

Sigmoid function should be applied on saliency map for conversion probability into [0, 1] range.

### Converted model¶

The converted model has the same parameters as the original model.

python3 <omz_dir>/tools/downloader/downloader.py --name <model_name>
python3 <omz_dir>/tools/downloader/converter.py --name <model_name>