Use case and High-level description

This model is an instance segmentation network for 80 classes of objects. It is a Mask-RCNN-like model with ResNet50 backbone, Feature Pyramid Networks block for feature maps refinement and relatively light segmentation head.




Metric Value
MS COCO val2017 box AP (max short side 320, max long side 480) 30.4%
MS COCO val2017 mask AP (max short side 320, max long side 480) 26.8%
MS COCO val2017 box AP (max height 320, max width 480) 29.8%
MS COCO val2017 mask AP (max height 320, max width 480) 26.3%
Max objects to detect 100
GFlops 56.433
MParams 44.920
Source framework PyTorch*

Average Precision (AP) is defined and measured according to standard MS COCO evaluation procedure.



  1. name: im_data , shape: [1x3x320x480] - An input image in the format [1xCxHxW]. The expected channel order is BGR.
  1. name: im_info, shape: [1x3] - Image information: processed image height, processed image width and processed image scale w.r.t. the original image resolution.


  1. name: classes, shape: [100, ] - Contiguous integer class ID for every detected object, '0' for background, i.e. no object.
  1. name: scores: shape: [100, ] - Detection confidence scores in range [0, 1] for every object.
  1. name: boxes, shape: [100, 4] - Bounding boxes around every detected objects in (top_left_x, top_left_y, bottom_right_x, bottom_right_y) format.
  1. name: raw_masks, shape: [100, 81, 28, 28] - Segmentation heatmaps for all classes for every output bounding box.

Legal Information

[*] Other names and brands may be claimed as the property of others.