Use case and High-level description

This model is an instance segmentation network for 80 classes of objects. It is a Mask R-CNN with ResNeXt101-32x8 backbone, PANet feature refiner with GroupNorm and DeformableConv operations and Adaptive Feature Pooling in all ROI-wise heads.




Metric Value
MS COCO val2017 box AP (max short side 800, max long side 1333) 45.36%
MS COCO val2017 mask AP (max short side 800, max long side 1333) 40.00%
MS COCO val2017 box AP (max height 800, max width 1333) 45.11%
MS COCO val2017 mask AP (max height 800, max width 1333) 39.84%
Max objects to detect 100
GFlops 899.568
MParams 174.568
Source framework PyTorch*

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



  1. name: im_data , shape: [1x3x800x1344] - 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.