ssd-resnet34-1200-onnx

Use Case and High-Level Description

The ssd-resnet-34-1200-onnx model is a multiscale SSD based on ResNet-34 backbone network intended to perform object detection. The model has been trained from the Common Objects in Context (COCO) image dataset. This model is pretrained in PyTorch* framework and converted to ONNX* format. For additional information refer to repository.

Specification

Metric Value
Type Detection
GFLOPs 433.411
MParams 20.058
Source framework PyTorch*

Accuracy

Metric Value
coco_precision 20.7198%
mAP 39.2752%

Input

Note that original model expects image in RGB format, converted model - in BGR format.

Original model

Image, shape - 1,3,1200,1200,, format is B,C,H,W where:

  • B - batch size
  • C - channel
  • H - height
  • W - width

Channel order is RGB.

Converted model

Image, shape - 1,3,1200,1200,, format is B,C,H,W where:

  • B - batch size
  • C - channel
  • H - height
  • W - width

Channel order is BGR.

Output

NOTE output format changes after Model Optimizer conversion. To find detailed explanation of changes, go to Model Optimizer development guide

Original model

  1. Classifier, name - labels, shape - 1,N, contains predicted classes for each detected bounding box. The model was trained on Microsoft* COCO dataset version with 80 categories of object.
  2. Probability, name - scores, shape - 1,N, contains confidence of each detected bounding boxes.
  3. Detection boxes, name - bboxes, shape - 1,N,4, contains detection boxes coordinates in format [y_min, x_min, y_max, x_max], where (x_min, y_min) are coordinates top left corner, (x_max, y_max) are coordinates right bottom corner. Coordinates are rescaled to input image size.

Converted model

  1. Classifier, shape - 1,200, contains predicted class ID for each detected bounding box. The model was trained on Microsoft* COCO dataset version with 80 categories of object.
  2. Probability, shape - 1,200, contains confidence of each detected bounding boxes.
  3. Detection boxes, shape - 1,200,4, contains detection boxes coordinates in format [y_min, x_min, y_max, x_max], where (x_min, y_min) are coordinates top left corner, (x_max, y_max) are coordinates right bottom corner. Coordinates are in normalized format, in range [0, 1].

Legal Information

The original model is distributed under the Apache License, Version 2.0. A copy of the license is provided in APACHE-2.0-MLPerf.txt.